ai-village-capture-the-flag-defcon31

[gaunt anchor]

@gusty warren you encouraged me to implement the idea I had yesterday , now I have three challenges in the queue 😄

[random minnow]

" you can easily catch up" this is by statistics.
p(num of trials to get correct | given easy task)

vs

p(num of people get correct | given difficult task)

[ornate marsh]

no pls

[glass bay]

...can you? won't any limitation that keeps the solution just reduce the search space and/or won't matter given the prior knowledge pre-buff?

[sand solstice]

kaggle has a CLI you can pip install, that’s actually how I’m submitting in my kernel

once you set it up with your API key it takes 2 secs

[gusty warren]

Yea, ususally they post the command in the competition page, so you don't need to guess the "competition name" for submission. I just didn't find it this time.

[gusty warren]

Havn't started granny (except uploaded an aplple and failed...), but from what I saw here, the brute-forcing one-pixcel attempt merely budged the probability. So what is the chance of there existing one magic pixel with the right number would change everything?... I mean, aren't there pooling layers in the CNN to exactly smooth the effect of one piexel?

[dense lodge]

did you solve 1 and 2?

[gusty warren]

no😂 ..hahahaha

I feel like granny 1 and 2 are those you can solve for sure, but just need some time. So trying other stuff so far

[naive umbra]

yay got sloth, cluster1 is next.

[ornate marsh]

i'm getting shower curtain, instead of granny smith. 👍

[past brook]

i think i will give up on mnist

and just focus on the other flags

[sturdy gorge]

yesssss, so happy, got cluster3

[gusty warren]

silly quesiton, for granny. Do we have to save the img first and load from disk like the starter notebook. Or there is in-memory solution...

[dense lodge]

you would need to send bytes that decoded are png or jpg so it's easier to just write a function that saves it

[gusty warren]

disk io is usually pretty slow, but may doesn't matter that much in this case?

[buoyant frigate]

"How are you doing?" "Searching flag in the taskbar instead of paint level"

[severe pasture]

you can use io.BytesIO to create a stream and save to that, much easier than saving to an actual file

[gusty warren]

Thanks so much! Learning little things along the competition

[sturdy gorge]

thats a very cool idea

[sturdy gorge]

1.5 flags away from bronze, this is so exciting

[past brook]

mnist has set me back 1week+

and i still havnt made any progress 😂

daily mnist rant

[sturdy gorge]

what is the general consensus on granny1 vs mnist, regarding difficulty?

[wind ether]

They're hard to compare, I found MNIST pretty easy but granny took me a while which doesn't seem like the consensus. The main difference is that granny gives you feedback on how well you're doing vs MNIST where you have to guess. Given that you prob will have to solve both for a medal I'd do whichever one you find more interesting

[dense lodge]

are we sure at least that granny uses MobileNetV2

and not like V1, V3 or ViT?

[jagged sluice]

granny uses specsavers

[wind ether]

Can't say, but moo mentioned that you don't actually have to match the model to solve. Probably why I found it much more difficult than others. I think it gives me an advantage in granny3 though.

[dense lodge]

sure but everyone who did claimed to have reproduced it

[jagged sluice]

???

[wind ether]

Not really, I'd say it's like 50-50. A lot of people brute-forced it

[gusty warren]

my wolf trying to pose as a granny and failed miserably, find it funny, lol....

[past brook]

wow

i got mnist

after all these years

[dense lodge]

I don't see how you can bruteforce it if it's not even clear what the task is. Clearly, it's not just submitting something that gets classified as an apple.

for any more sofisticated ranking of classes you might want to achieve you need to have a proper adverserial attack

[past brook]

i really dont understand the hint

on mnist, after solving it

[wind ether]

Yeah I didn't either

[past brook]

really threw me off not gonna lie

[gusty warren]

oh, I was just having fun and tried to build a naive model before reading any paper, it is planned to fail. But it's FUN and may help me understand when I do read the paper later

[gusty warren]

Congrats! But do you understand the hint after you solve it? Like, for IP 1&2, I understand nothing even after solving it.

[past brook]

i kind of get it, but it threw me off more than it helped

[wind ether]

I've been trying to understand it to see if figuring it out helps me think through CIFAR's hint but no luck

[past brook]

i think i might know how to solve cifar

should i mess up my sleep and try it now or do it tomorrow instead of working 🤔

[severe pasture]

nah I didn't reproduce it

just queried against the server

granny3 getting 'close'

[past brook]

hmm maybe i didnt know the answer

[hollow coyote]

lol, i've been trying to manipulate that test_noise.wav this whole time. DOH

[random minnow]

"h, I was just having fun and tried to build a naive model before reading any paper,"
the recommended steps are:

1. read last year compeition solution
2. ask google and github,kaggle
   3.read paper

being "simpler problem", it means that quite a number of people has the solution. then we ask, how did they get the solution?

either, it is simple and they can create it from their mind after a few trials, or they could have found the solution from somewhere.

when you face a problem not only think of the solution, also think of how others have found the solution, how likely is others already solved the solution ... ... look yourslves as part of the statistics (chatgpt also proved that "part of intelligent" is atcually memorization and probability)

"Can't say, but moo mentioned that you don't actually have to match the model to solve."
his exact words are "Some people did it this way - you don't need to do this. In our solution for testing, we don't use a surrogate model."

[gusty warren]

Exactly, granny 1 is not a trivial one, althoug it seems to be easy at first glance. My feeling is that, granny series seem to be challenging (especially if you aim for granny 3), even though there are a lot resources already. So maybe it helps to think of the problem before reading solutions, given the context that the whole competition wants you to think outside the box...

I was thinking of an unified approach to solve granny 123 together. I think it could work, but most likely can't be implemented within the competition time frame. So better to forget about it and start reading papers.

[mild shale]

does the clue given by the organizers really helps us

[grave frigate]

Do we really need 1 in granny1?

[brave briar]

hehe it remembers me last year on hotterdog 🙂

I learnt a lot, like that to fool a machine, one has to watch as a machine not a human !

[abstract rose]

Inversion still cryptic after many attempts ... either I'm totally on the wrong way or what the spirit returns is the input for another puzzle

[brave briar]

I am at the same point ...

[gaunt anchor]

most of us (most likely) are at "the" same point :/

[timid steppe]

For the hush challenge, do we need sound processing knowledge to solve it?

[craggy beacon]

probably, the aim of competition is to attract our attention to the new areas of knowledge

[pallid trench]

Hi, any hint on how to solve pixelated..

[craggy beacon]

try harder

[pallid trench]

Last 5 days. Stuck.. With Jenny

[abstract rose]

Jenny is another joke to distract you

[pallid trench]

I am trying brute force, it is not working

[gaunt anchor]

You should try and try .... I stucked for a week 🙂

[amber sapphire]

Mnist😭😭😭 wasting time with no solution...

[craggy beacon]

cool, but don't share working approaches it is prohibited

[fervent obsidian]

You probably shouldn't say much about how you did it

[stoic shale]

ok sorry guys, I have deleted it

[craggy beacon]

here we can share only pain

[worldly zinc]

working on pixelated for many days too

[pallid trench]

Is pixelated harder than granny 1

[minor falcon]

depends for who

[random minnow]

i didn't know posion AI can be put to good(?) use

this tool add some "hidden data" to your image, so that if people used it to train their model (like diffusion model), it will posion their training

[craggy beacon]

so it is classified as generated by special model

watermarking of some kind

[random minnow]

not watermark, it posion training

[acoustic temple]

I think it is kind of attack, right ?

I mean data poisoning is kind of attack, but int this method, they use it to protect model

but what is its role in this competition 😄

[craggy beacon]

oh yeah you are right, that should be poisoning of datasets by misslabeling them but " Nightshade changes the pixels of a digital image"

[random minnow]

no role, just general knowledge related AI poisoning

[icy notch]

will it have any impact if only 1 sample from million is poisoned?

[random minnow]

"When they fed Stable Diffusion just 50 poisoned images of dogs and then prompted it to create images of dogs itself, the output started looking weird—creatures with too many limbs and cartoonish faces. With 300 poisoned samples, an attacker can manipulate Stable Diffusion to generate images of dogs to look like cats. "

300 images

the paper: https://arxiv.org/abs/2310.13828

maybe it will becoome next year kaggle CFT problem task

[craggy beacon]

cool stuff because I thought big corpo would remove artists names from labels so there will be no way to associate generated images with artists

[acoustic temple]

until next year researchers will find a way to deceted and clean poisoned images,!

[minor falcon]

good luck to detect a one pixel attack 😅

[craggy beacon]

i don't think it will survive granny2 protection

[final path]

i think 1px attack could be easily detected, because usually this pixel values should be very far from surrounding pixels and it's visible to human eye

[acoustic temple]

agree, there are even studies around it: https://www.hindawi.com/journals/wcmc/2021/8891204/

[minor falcon]

yes but you might also induce a lot of false positive i guess if you try to simply look at distance between two pixels no ?

[final path]

yeah, there's classic precision/recall tradeoff. but fgsm attack should be harder to detect because of very small differences

[random minnow]

one has to note that one pixel attack is not always 100% possible. (It is unlike granny1, where is there is always a solution)

one can google for a benckmark/leadboard ranking that rank all papers for image attack . the sucess rate for imagenet is actually not very high

[sand solstice]

sigh rly close on pixelated but im having trouble with the last bit

[minor falcon]

from what i was reading, while 1px attack is not always possible, "few pixels" attack considerably increases the probability of success (3-5px). Also we don't necessarly need to poison a full dataset to protect it I guess, poisening 10-20% of the data might be largely enough

[sand solstice]

wish the ocr was better

[random minnow]

"wish the ocr was better". There is a chinese saying 山不转路转，路不转人转，人不转心转: " If the mountain won't move, build a road around it. If the road won't turn, change your path."

carefully analyse the whole flow and system and see what/how/who/which/when can be changed and what cannot

[final path]

yeah, success rates for 1px attack on imagenet are pretty far from 100% as I see from papers. targeted attack make things harder, and 768x768 image which is resized makes it even more harder. 🤔

[sand solstice]

if this helps me unblock I will cite this in my writeup lol

[glass bay]

did you all managed to 100% match the server model locally? i've got only up to 4-5 digits and it seemed not enough

on granny*

[acoustic temple]

tried 3 different model, even coulnd't find the match model, but as I heard moohax mention that one don't need to match model

[minor falcon]

@glass bay this message from moo might interest you

[glass bay]

for granny 1-2 most definitely that's true, but for granny 3 idk tbh

[minor falcon]

due to amount of queries limitation ?

[glass bay]

yes, and due to other attacks being simpler to perform in comparison to onepixel

[brave briar]

but moo was referring to all granny levels or only granny 1 ?

[glass bay]

it is unclear from the context it seems

question is: is granny 2 model the same as granny 3?

[thorn quarry]

one confirmation do we get flag if we get 0.33333 for pos, neg and neutral right?

[brave briar]

hehe you can test it simply 🙂

[fervent obsidian]

People posted their results getting all 0.3333 but no flag

[glass bay]

true

[thorn quarry]

also moo once said find out how passwords work i didnt find any relation between passwords and sentiment. any suggestions?

[sturdy gorge]

I remember seeing in this chat something about the 0.3333 differing in like the 6th/7th decimal case

[carmine stream]

I get the same score as the original sentence, but I don't get the flag, maybe there are more hidden decimal places in the system?

[glass bay]

... or something entirely another you need to account for except scores

[sturdy gorge]

patiently waiting for workhours to end to try granny1

[gaunt anchor]

hush server is slow again ? anyone tried ?

@limber flower @olive ledge can you guys please check if the server is down or slow ?

[sturdy gorge]

server extremly slow for semantle2 as well

[brave briar]

can we submit regex for inversion 🙂 ?

[random minnow]

old competition for those who are interested

[violet trellis]

i got 0.995 in granny 1 and 2 with no model, though no flag 👀...

[boreal spear]

if you submit an apple obviously no flag for you

[versed jungle]

right now, all the sentences I submit to passphrase give the same score ...

[olive ledge]

Looking.

[half plinth]

you are probably sending fixed benchmark string all the time

[versed jungle]

a yes indeed !

[gaunt anchor]

oh my god

I solved inversion

[abstract rose]

Spirits are with you, deamons with me. Congratulations

[gaunt anchor]

23 flags , I didn't catch gold yet :/

[minor falcon]

i share your pain

[sturdy gorge]

23 not enough for gold now?

[gaunt anchor]

Again same problem , I had the key .... I lost the door 😦 😦

wasted days daaaaays

[abstract rose]

How long to solve it?

[sturdy gorge]

wont you get gold when they remove the deleted account from LB?

[brave briar]

Maybe it can t be revealed, but when you got the static flag, it was evident you had it or you did lot of try and retry ?

[gaunt anchor]

I can't say more :/

but its a good feeling I hope you guys feel ... when solving it i mean

[gaunt anchor]

No currently I am at top silver :/

[queen garden]

The dopamine rush when solving a problem is what we're all addicted to 🙂

[fallow valve]

No, you are gold, you forgot the deleted account

[wanton patrol]

I think the 4 guys who submitted within an hour of each other took the last golds yesterday

[acoustic temple]

this was the answer I was searching, when I'm questionin myself, why I don't eat, sleep and just want to solve more questions

[brave briar]

no problem, I will reformulate my question : "how great do you find the problem and the solution to solve it at end ?". Let's hope we get this feeling too :). I feel I am close but still can't interpret correctly the spirit message !

[queen garden]

@brave briar stop trying to get people to tell hints

[gaunt anchor]

Hmm the deleted account doesn't count so .... I think I am last at gold currently

[acoustic temple]

proud of you my NameMate 😉

[half plinth]

my dopamine rush stopped on the 20-ish flag. The remaining tasks without a stated goal where you need just try and try make my crazy

[craggy beacon]

same, and the inversion challenge kinda kill it for me, too much variables even with last year solution with emnist dataset

[thorn quarry]

Guys one doubt in pass phrase my scores difference is 0.002 in negetive and hence the same in positive so anything is there a logical way to optimise it or have to check for a whole different method instead

[gusty warren]

Did you solve Semantle 1 & 2? I believe there are many (logical) ways to do these type of optimization

[random minnow]

" I believe there are many (logical) ways to do these type of optimization"
ask google

[boreal spear]

brute force is logical
guess is also logical

[gusty warren]

Reading other people's logical ways after the competition will be fun

My granny is taking around 4s per request... anyone having same issue?

[gusty warren]

systematic guessing saves you time though

[sturdy gorge]

semantl2 at 0.95, still very far away

[random minnow]

even Autonomous vehicles. can be attacked

i suddenly think athat all ML/AI are unsafe

[craggy beacon]

electricity is unsafe

[gusty warren]

I think the spoofing lane-changing... it's not just an attack on AI, it's an plain attack...

[empty bane]

you can do the same attack against human drivers... just pretend to be about to turn into them and they'll brake too

[random minnow]

hmm

with masked pretraining becomming popular (like gpt and mae(masked auto encoder)), i think we are going to have a huge problem. basically, in generative pretraining, you are predicting the data itself. hence it is asy to do train data inversion. (e.g. https://gradvit.github.io/)
my friend once told me a joke: waht is the difference between software and ML enginner?
[daily work of software engineer]:

1. write code
2. make a bug
   3.debug
   4.repeat

[daily work of ML/AI engineer]:

1. make an new ai aglorithm
2. discover flaw
3. make another ai aglorithm to cover up the flaw
   4.repeat

[grave frigate]

I'm doing granny1
My score is almost 0.999 on local
But on api is hardly improved by 0.0001 why?

Any help?😅

[thorn quarry]

Ok I have to follow the approach which we used in semantle2 after getting close to the score. Am I correct? @gusty warren @random minnow

[grave frigate]

I'm so done😭

[queen garden]

@random minnow People have already told you but you're sharing way too much. Here you're spoon-feeding information - that is not in the spirit of the competition.

For people to actually learn from the competition they need to try to understand what is going on by themselves, not wait for someone to give them the answer...

[random minnow]

i didn't given them the answer. i only tell him the direction to find their own answer

[craggy beacon]

i think it is ok because tasks can be done without understanding underlying theory

[grave frigate]

++

I'm just asking for some direction
I have done this much work🥲

@random minnow apologies brother
I won't ask anything further

[wind ether]

I think it's time to give up on inversion, gonna save it for last if I get that far

[craggy beacon]

i decided to give it a break after i saw this board

[wind ether]

My main issue with it is that it's just so subjective--I love the problem, it's extremely interesting in itself and I learned a ton about model inversion in general, but I'm just not able to find a solution from the 5/8 pieces of data that we have and trying to brute-force the last 3/8 hasn't been helpful. Just wasting a ton of time on my end trying to find a suitable word from the combinations of letters/numbers that are most likely to match at this point.

[lost relic]

These days I frequently have this quote on my mind 🙂

[gusty warren]

Lol... so should we start trying French characters?.....

How does that board even work. At the bottom there is goodbye in french?

[gusty warren]

I havn't solved granny1, but it doesn't seem like the objective is to get a perfect 0.9999 score, that makes no sense as "adversial attack" (last year's hot dog was weird)

[grave frigate]

My issue is with being unable to increase any substantial value for api result

[rocky jacinth]

At long last got cluster3. Exhausted.

[wind ether]

Maybe the API result doesn't really matter at all beyond beating out wolf?

[grave frigate]

Finally i'm dangerous

No flag though

[gusty warren]

We all get a teapot troll as reward of being dangerous..

[grave frigate]

😂😂😂

[gusty warren]

Did someone just stop spamming the server... my granny request is getting way more responsive

what if.... you reach a certain spot on the leaderboard and don't feel like solving more problems, and start to DDoS the server...

[grave frigate]

do I need to be more dangerous or what?

[final path]

literally me solving CIFAR

[gusty warren]

I wish I know.... I keep looking at the example {"thing": [1, 2, "here"]} and think maybe there is a static flag hidden in some array, lol.....

[ornate marsh]

cluster 3 was one of my favorites!

[rocky jacinth]

@ornate marsh I found it very demanding, and needing very precise work.

[sand solstice]

same sadly

[median mauve]

granny slow :/

[dense lodge]

I was trying to match the model yday for hours without much success, now opened my notebook and did something turns out my model perfectly matches api oO

[sand solstice]

I’ve played around with pre/post processing, but sadly everything I have tried, while decreasing local/server error, does not get me anywhere close

Attacking a slow black box model without the gradient is, shall we say, painful

If I was rational I would throw in the towel and hope for bronze, but alas sunk cost

Probably something to do with not being able to judge how close I am to the remaining flags

[sand solstice]

real

[gusty warren]

I was hopping to solve as black-box, but today's response time.... kind of forcing us to match the model

[ornate marsh]

timber wolf is watching

[glass bay]

say one thing - is it non-Imagenet mean/std normalization that we'll have to grind up to 5th decimal place in hopes of perfect coincidence?

[violet trellis]

Well, you gave me some hope :).

[dense lodge]

ah yes, finally silver!

[naive umbra]

Solved my last cluster challenge, cluster1. Hardest part was reading the model correctly. Luckily I stumbled upon the Kaggle discussion board.

[dense lodge]

@gaunt anchor which one was your 23rd, if that's not a secret?

[gaunt anchor]

I already mentioned it ... inversion

[dense lodge]

sure but I can't seem to be able to search by name :/

[gaunt anchor]

now the big four players (granny4, CIFAR , passphrase and hush) , I already have plans to them ... but I think its going to be loong loong two weeks

[dense lodge]

do you think having 22 now will be enough for silver? I'm like 36th

[gaunt anchor]

You never know ... aim for gold so if you got silver you still be happy 🙂

[dense lodge]

ofc but I do have other stuff going on and don't wanna waste too much time on this 😄

[gaunt anchor]

I need atleast silver to push more to master ... anything more is welcomed 🙂

[dense lodge]

that's what I'm going for as well

[gaunt anchor]

solo gold is more great ! ... money is greater hhhhh

[dense lodge]

I have solo gold already so not a big deal for me

[gaunt anchor]

Ya , well I need it so this is my chance to catch it (I hope)

[dense lodge]

good luck 🤞

[gaunt anchor]

you too 🙂

[nimble matrix]

Does pickle need in-depth knowledge of pickling and unpickling?

[gusty warren]

flag or not, we begin with no in-depth knowledge and end up with in-depth knowledge...

I heard somebody said that the prior knowledge of pickle set him back rather than help him. Don't know if that's individual case or general

[nimble matrix]

Yeah I have been doing pickle since the ages and I don't think I am any where near the flag

Meanwhile I got sloth, semantic 2 and MNIST
But pickle is driving me crazy

I think I should move to the granny series

[gusty warren]

pickle is annoying because there is no feedback at all

the only feedback you get is a troll

[nimble matrix]

Yes that's what's annoying me

[nimble matrix]

Initially I didn't know that it was trolling me until I asked here

[hard perch]

Do we need to know about DNS/IPv6 stuffs for IPs or they are just like witfs?

[nimble matrix]

Go through the old messages
I remember someone has answered this

[jagged sluice]

¯_(ツ)_/¯

[nimble matrix]

Hehe missing this

[topaz ember]

I don’t know how others did pickle, but even after getting a flag I have no clue about the overall idea behind this challenge.

[grave frigate]

Has there been any clue regarding pickle
Like any?

[topaz ember]

I saw some people in this channel recommended to re-read the prompt, waiting for the end of this comp to see how it helps. May be there are several different solutions, mine was not related to the prompt text at all.

[twin wren]

0.99 GRANNY2 NO FLAG

[gaunt anchor]

granny3 : 0.0006914532859809697 😔

[twin wren]

just want to know

has granny 3 solved by anyone?

801-867-5309

[gaunt anchor]

as far as I know ... not yet ... but maybe the 24 guys did it !

[twin wren]

Is it hard to solve pixelated without knowing jenny's joke

[random minnow]

801-867-5309 : this is a real phone number. call to get the flag

[gusty warren]

Are you talking about granny 1 or 2? matching granny 1 seems quite straight forward

[twin wren]

i dont even know who is jenny lol..

need cultural reference to solve a ML problem..

[jagged sluice]

¯_(ツ)_/¯

[sand solstice]

I step away for 1 hr, get 2 flags, and im back in silver 🤦‍♂️

time for a break

[twin wren]

[sand solstice]

overthinking is no joke

theres a life lesson in there…

[timid steppe]

Same here, I get the flag just by chance

[gusty warren]

sounds like an LLM?

[grave frigate]

You too stuck in pickle?

[gusty warren]

yea, I kind of give up on pickle. Switching to granny

I am not learning anything doing pickle, unlike granny. so much fun

[twin wren]

So as MNIST

🤷‍♂️

[gusty warren]

trained with the granny 2 api, and validated with granny 1 api. And was wondering why the score not going up..... just funny bugs everywhere

[grave frigate]

I too ditched granny🥲

[random minnow]

l look through previous competitions. actaully before last year DEF CON 30 at kaggle, there are also many previous task realted to image ML/AI. you can take a look at (past events DEFCON-26,27,28,...). many nice articles on their https://aivillage.org/blog/
https://aivillage.org/events

[jagged sluice]

Saturday talks were good too smh

[gaunt anchor]

Granny3 I was able to reach 0.0006924299523234367 .... hmmm

[random minnow]

your slides?

"Granny3 I was able to reach 0.0006924299523234367 " i have strong feeling that granny3 should be one of the unsolvable tasks at the end of this competition

i think we need at least 100 pixels. 1pixel seems mission impossible

[gaunt anchor]

I am now at 0.0006927x .... I will give my current strategy some time to see

[buoyant frigate]

Maybe a little less than 100 if something like p(c)=0.001(-epsilon) if c!=granny smith and p(c)=0.001+(epsilon) if c=granny smith is accepted

[fervent obsidian]

I got 0.0006927589 as well but then it didn't get better from there. I'm now re-evaluating my strategy.

[gaunt anchor]

0.0006927594658918679 the value I am stuck at hhhhhh

[buoyant frigate]

you just need to reach 0.00101 and reduce all the rest under it, 0.00069 is just 69% of the minimum score to have granny smith as the classified class. (the problem is reaching a condition where 0.001 is enough to make it be the winning class)

[queen garden]

Could get 0.000749 locally. However achieving that score with the API is another thing ...

[buoyant frigate]

(hope what I wrote above is not treated as hint) I'm far from 0.000692 locally, let alone solving the problem

[queen garden]

Not a hint but a boost for my motivation 🙂

[acoustic temple]

I'm stuck at Granny1-2, and conversations are just around granny3, me like 👀

[fervent obsidian]

For granny2, I'm sending the edited timber wolf picture the api returns granny smith but doesn't give the flag

[buoyant frigate]

I'm glad, motivation is a great deal in these challenges.
Now probably I should go back into minimizing the top predicted class score, just to see what's my limit with one pixel

[acoustic temple]

same

[buoyant frigate]

One last message before going.
Let's try the approach of Phoenix wright game in the first case, and rephrease it a bit.
"Don't try to understand if your assumption is right, go on thinking your assumption is right and try the exploits that your assumpion suggestes"
And always change your point of view if the problem as you frame it seems unsolvable, maybe that's not the real problem

[minor falcon]

i think i know exactly which pixel you are at :p

[twin wren]

Pickle...

i am so sure it is very dangerous

[violet trellis]

A lot of teapots around...

[buoyant frigate]

Http status 418?

[violet trellis]

Yup

[hard perch]

@lunar blaze asked it, but before getting any reply, he said he solved it ;-;

[buoyant frigate]

yeah, I tried so many things in pickle that in the last version of the notebook I saved I have an XSS embedded in the page and i alert(1) when opening the notebook.
Maybe I overdid something. For sure I'm overthinking

Why can't I just
"delete from BRAIN where topic='pickle'"
And start from 0?

I did pickle. Well I deleted enough from my brain

[grave frigate]

And here I'm
Still stuck

[abstract rose]

Endless loop on Inversion, I'm sure I've it but I don't have. Do we need to smoke cannabis to get it?

[brave briar]

Yep at the same point, here. I am pretty sure I have all the puzzle pieces. But it does not match...

[minor falcon]

We are 3

[jagged sluice]

Hehehehaw

[twin wren]

waving keyboard for 2 hours already

no miracle

[glass bay]

Inversion hint: it's trying to tell you something

[boreal spear]

it's telling me to give up

[outer sundial]

But 4,5,7 are not ready to tell anything

[minor falcon]

i got them telling me something, but something wrong obviously 😅

now i'm just trying about 1M400 comb and #yolo

[brave briar]

ouch you are ambitious, I generally limit to some thousands 🙂

[minor falcon]

yeah well lets say i gave up for now so my stuff run and if my miracle i find something, good news, otherwise, too bad! :p

actually I just hope i'm far off the actual solution, because I will be kind of pissed at myself if I miss the answer by 1 character somewhere

[final path]

maybe this will help

[severe pasture]

aaaaaa inversion finally!

only took two weeks

[minor falcon]

i remember seeing that one, not ringing any bell for now ! but thanks for reposting

[craggy beacon]

do u like your solution?

[severe pasture]

a little

[abstract rose]

Yes, and I think I understand what but then I'm always wrong at the end

[craggy beacon]

ehh I thought we can actually communicate with model by asking questions😅

[glass bay]

But... It's a single model with 8 outputs...

[amber totem]

Yep but don’t know if treshold is really low or if it’s just not the right way for 457

[craggy beacon]

yep there are not that much reasons for that

[glass bay]

Alright, I tried to nudge y'all in the right direction

[craggy beacon]

maybe i am using wrong datasets as moo said or wrong dictionaries

[hard perch]

Going mad with the IP ones ;-;

[twin wren]

I need pickle to give me silver 😡
Won’t sleep until I get pickle

[hard perch]

Dangerous pickle detected

[stoic shale]

give me the flag :))

[random minnow]

"Maybe a little less than 100 if something like p(c)=0.001(-epsilon)"
by using local model, i can compute/probe the receptive field of one pixel. i also have the timber wolf image. i can compute the max chnage in feature value and logit due to the largest change in single pixel value .... conclusion .... maybe not possible?

[ornate marsh]

https://tenor.com/view/timcook-chequered-flag-f1-apple-wave-gif-26980817

[wanton patrol]

just a reminder that discussing the conclusions you've reached during your research, however obvious to you now, are shortcuts for many people

I can see that there is a dire (wolf 😆 ) need for a team competition of this kind 😄

[acoustic temple]

what is this now, I'm getting granny as first prediction in Granny2, but no flag !!! 🥲

[gusty warren]

but did you get the flag for granny 1?

[dense lodge]

It also needs to be high probability it's not enough to be like 5% granny top pick

[devout jasper]

now 8th day in a row trying to solve granny 1

[valid cobalt]

Oh, that's not what I was supposed to do :((

[jagged sluice]

maybe you need .3333 .3333 .3333

[random minnow]

i have few suggestions for next year kaggle CFT (after reading some interesing ml attack papers)
1: guess the prompt, given AI generated image (e.g. from stable diffisuion)
2. instead of the prompt above, guess the style or other attributes of some given prompt template.
3. uncover the reference given intput prompt and ouput generated results (e.g. if we have input prompt and ai generated summary, can you uncover the source documents where the summary is made from)

[ember relic]

my god

i just spent 5 hours finding a mistake in cluster 3

sometimes sleeping truly is the best option

[twin wren]

Found the only logical shape for MNIST

and you tell me its wrong?

FK

[gusty warren]

welcome to the club

[gusty warren]

I am starting to wonder whether granny 3 is easier than MNIST

[twin wren]

at least its something u can optimise

[random minnow]

i find a solution for mr wolf to hide himself: https://youtube.com/shorts/IAh_GoJL_GE?si=HTq6w-mj7aFbxQcQ

[gusty warren]

I got a feeling that quite a few people solved granny 3 already, just being silent about it. I mean, if the task is really just changing one pixel.

[twin wren]

LOL

finally

this is the way

[grave frigate]

So what do we even count😂

[lyric token]

wasted time

[grave frigate]

It's , for me, uncountable infinity

[grave frigate]

Is everyone collaborating in background or what
How are these scores moving together 😶

One reaches 21
And everyone reaches 20

[abstract rose]

@olive ledge About the Inversion prompt 'When you believe you have the static flag, submit it here', the static flag is not returned by the API correct?

[glass bay]

yes

[craggy beacon]

I hope next challenge involving letters will be using some rare language to remove a bias toward native english speakers

[naive umbra]

Nice got IP1 and IP2 in 5 minutes. Less luck with pickle so far.

[acoustic temple]

are you comparing pickly with IP1-2, do you mean there is a LLM behind it ?

[naive umbra]

I don't know yet. At least I have a dangerous pickle but I'm still figuring out if that is the way to go.

[twin wren]

Yes…
I don’t know anything about janny joke

[inland mural]

I took a week off and worked on sloth again and solved it in 5 minutes!
Rest is important

[grave frigate]

Hints😭I need hints

Besides resting & not overthinking

[twin wren]

The put MNIST and pickle to keep people away from medal

😭

[craggy beacon]

google it

[hard perch]

This IPs are killing meee

[lyric token]

IP and Pickle may google. MNIST - I don't think so

[random minnow]

"MNIST - I don't think so"
are you sure? don't assume. try google it

[craggy beacon]

i think CIFAR is googlable

[random minnow]

it is the prompt you enter in google that decide you get the answer or not

[lyric token]

I did MNIST, i think it is most useless task in this challenge

[craggy beacon]

inverse is the worst

[lyric token]

maybe, I'm still trying to get flag from trash predictions))

[random minnow]

Spark of Genius = adversial noise

google search can give you lots of noise

https://www.nationalgeographic.com/science/article/eureka-insight-newton-archimedes-genius-science

this is why newton "discover gravity when lying delow apple tree"

[violet trellis]

I think it's just he wasn't hungry that time. Otherwise, he would have eaten the apple and slept.

[lyric token]

Has anyone been able to get good outputs for 4,5,7 in inversion?

[craggy beacon]

for me a bigger problem the amount of different chars with pretty confident 0.999 predictions for each place, 10-15+ letters for each pos

[minor falcon]

thats the question that is haunting all of us still stuck on inversion

[lyric token]

I even got totally different shapes and back grounds for the same positions with perfect 1.0 score

[brave briar]

Hum oddly I am stuck elsewhere ...

[glass bay]

pretty sure there was something on these sockets and they were massively unlearned to obfuscate whatever there is

[abstract rose]

Unlearned or never learned

[glass bay]

they definitely dont want an ai to tell you what it wants

[nimble matrix]

"dangerous pickle detected" is it just to troll me or what?

[modern anchor]

https://tenor.com/view/confused-pig-gif-8664053796761769219

Same with me. Dunno what this is

[carmine stream]

Me too, I don’t know if it’s a reminder or a confusion

[abstract rose]

IMO It means that you've been noticed by the police

[nimble matrix]

I am in Jail only since I started pickle

[thorny ivy]

First time doing a competition like this, I'm pretty lost I like it hahhaah

[thorny ivy]

Kind of, FBI is in my door, but I'm continue trying pickle before to open

[gusty warren]

But I sent bs byte string that makes no sense to it and it still told me dangerous pickle detected...

granny=0.22, wolf=0.14, still no flag for granny 2..... wondering whether the granny prob is not big enough or the generated image is too deviated....

[random minnow]

since you already pass granny1, why don't you test it with granny1 (lower garnny probability, deviation, etc)?

[final path]

i think discussions like this about granny are very hint-ish 🙂 because figuring out what to do is a major part of a task too, I personally spent like 5 days digging the wrong way

[gusty warren]

Because I am not sure the different levels share the same threshold? but yes, I should test it anyway....

[random minnow]

you should think of it as: if you don't know, then others would not know too. but others managed to do it. so it is a trial and errors. hacking itself is trial and error.

it would be faster if you try it yourself than asking here

usually you may think that people finished the task they may have discover secret you don't know. but in reality, maybe everyone is seeing the same thing. just that they try more and has less bug

[final path]

yep, but with all that info you could try less with higher success rate 🙂

[random minnow]

this is the reason why in the end, the lb score is the same (becuase everyone sees the same thing). just that some people achieve faster.

[final path]

e.g. you can spend like week figuring out what to do in granny if you do it from the start, but in the middle of comp you can use all public info and spend like 1 day instead. and invest this spare time in other stuff

[minor falcon]

oh i got a 0.000693

[grave frigate]

I've the same score for granny on granny1

Samesies bro😂

[minor falcon]

might be due to rounding actually... I print with 6 digits... So it might still be the 0.0006927

[steep nexus]

Why is the OCR choking so hard on longer inputs ? It's driving me so mad 🥲

[gusty warren]

The probability of the OCR getting the string right: (acc% per char)^(len of string)...

[glass bay]

imo it's worse than that because too frequently i've had my input's last character just forgotten

[empty bane]

it definitely gets worse when its longer

[random minnow]

it is interesing for ocr becuase this technology has been around for many many years.
even before deep learning or advanced computer vision.

e.g. to scan car license plate, to read bank check serial number

[steep nexus]

Like, the exact same input will be read perfectly when split but reading the whole thing is impossible... makes no sense to me

[random minnow]

so people still manged to get it work desipte the flaw in technology

and that is the trick to make AI product. An sucessful AI/ML product(in which model is only one small part of it) is one that works even it the algorirthm has flaw

[final path]

passpharase is driving me crazy. I'm about to draw a diagram with relationships between sauce, sentiment, model, chef and Bitville 😅

[minor falcon]

what's is even driving crazier about passphrase is that apparently taking a shower + 3hrs of work is enough to get it right :p

[final path]

maybe 3hrs of shower will do? 😄

[queen garden]

20+h of work on passphrase and I don't feel like I'm any closer to understanding what I have to do 🙃

I never shower maybe that's the issue

[final path]

At least we could get differencedifferent responses from API, unlike CIFAR

[minor falcon]

you don't like this "try again!"

actually you can get also another message "wrong input shape etc..."

[final path]

Simple "try again" challenge

[minor falcon]

inversion is only "try again!", that i saw almost 1 million time i think as of now

[random minnow]

"I'm any closer to understanding what I have to do"
google can help

[abstract rose]

I've solved granny2 after 20min cycling

Tomorrow cycling again, I hope it will help for Inversion 🙂

[gaunt anchor]

CIFAR evertime I think I got a logical input ... it gives me the two words that everybody hates currently "try again" !

and thats the same for Passphrase

hush gives anything ! there is always numbers ....
granny smith , I add a note that once I am at the supermarket I will change it price to 0.000692

[minor falcon]

round your forecast, it will give you the feelink of having 0.000693, or even more rounded 0.0007. Or even more rounded 0..001

[ornate marsh]

same but with MNIST 😄

[random minnow]

you see many "advml" appearing in the submission url. i think i finally undersatnd what this is

[toxic willow]

hi- i just started this challenge. is there a particular format to submit the answer for cluster 1? the response i get is {s:#.####... }. but the submittal format is {cluster1:"empty", etc.}. i've submitted the test flag and the value for cluster1 response, but my public score is 0.

[terse moss]

When you get teh correct answer, you will know it prior to submitting.

[past brook]

you should create a seperate notebook where you submit the flags .
https://www.kaggle.com/code/kononenko/ctf-submission

something like that

[terse moss]

If you submit the Test flag correctly, you will get a point.

[past brook]

and just copy paste the correct flag when you get it into the dictionary

[terse moss]

I was working on Hush for so long; I eventually got the 'I don't like the sound of that' message and thought I had made a breakthru...

[gaunt anchor]

I was working on it for 10 days so far …

[earnest sparrow]

Is there any tracker for which question have not been solved yet...i see 24 solved what are the remaining 3?

[gaunt anchor]

Most likely hush , granny3 , passphrase and CIFAR is what left for the guys with 23 flags .... (most likely)

[craggy beacon]

CIFAR has been solved by decent amount of people as moo said 10 days ago

[earnest sparrow]

Its fine i get the idea 😅

[gaunt anchor]

if CIFAR has been solved that much then we have to hurry more to keep our ranks ! 😅

[craggy beacon]

passphrase has been solved too

[gaunt anchor]

ya once as far as we know ! I think

[earnest sparrow]

either way i am noway near ...i am just trying to understand the riddles

[final path]

but idk how cifar can be solved, because no one in this chat claimed this..

[jagged sluice]

cifar is easy

all you have to do is send the entire cifar dataset to moo's address

printed or faxed

then wait for flag to appear in your mailbox

[limber flower]

Preferable printed, the price of paper is insane these days

[jagged sluice]

gotta use gloss paper too

otherwise you might only get half a flag

1 image per page

[modern anchor]

Finally solved MNIST after so long. Disgusting!!

Wasted yet another 3 hours on CIFAR 😭

[jagged sluice]

3 hours? rookie numbers

[violet trellis]

Well for me, i wasted almost a week on pickle and pixelated. And finally I solved

.
.
Neither

[thorny ivy]

I just have a lucky shot and solver MNIST in around 10 mins, but I'm crying with CIFAR, no clue at all

[hard perch]

Finally got my email sent
Damn those IPs

[hollow coyote]

Bah, so close to cluster 3

[hard perch]

Same. Frustrating

[acoustic temple]

Did you understand the 2nd part of the clue to solve it?

[tepid zenith]

I am trying to make it like the sound of almost everything 😦

[twin wren]

accidently delete all my file when doing pickle...

emotional damage

[hollow coyote]

Finally got it lol

[thorn quarry]

Can anyone give me direction to solve MNIST , I tried to use the frequencies of pixel values at the image index of 255,23435 but no flag,i'm banging my head for three whole days and also did a 3 hr shower but no luck so far 🥲

[thorn quarry]

Is it due to wrong data set? 🤔

[hardy barn]

Hello everyone. In the "Pixelated" task, is it required to send an image with a special password/word, or should I modify the original image?

[glass bay]

Whatever gives you admin privileges

[twin wren]

brute forcing phone number

😤

[grave frigate]

😂😂😂

[grave frigate]

The most hint you'd get here is "count"

[gusty warren]

yet the only number I can count correctly is the hours wasted on MNIST

[grave frigate]

How do you express infinity in python?

[gusty warren]

something like np.inf?

[grave frigate]

Joke...ruined in an instant

[carmine stream]

Why are my local MobileNetV2 results inconsistent with the api results? Very strange

[twin wren]

spent days to figure out why

[ember relic]

is it possible to automate things in pixelated since the ocr is pretty bad?

[gusty warren]

I believe that's most time consumeing part of pixelated... lol

[ember relic]

hmm i was hoping not to automate it at all and manually change stuff

but i dont seem to be getting anywhere

[gusty warren]

I mannually changed stuff to do proof of concept, to see whether it's worht the time to automate and tune the img

[jagged sluice]

🛌

[naive umbra]

For granny I get 99% with my local model but the API still thinks it’s a wolf I hope I’m on the right track tho

[carmine stream]

I've been stuck here for a long time

[gusty warren]

I hope it's a joke...lol

[twin wren]

For Granny, at least you have something to deduce and progress

for pickle and MNIST

just constant guessing

[hardy barn]

MNIST is real easy problem. Just think about what you can count

[outer sundial]

You should delete it

[twin wren]

i thought that is very trivial ..

[gusty warren]

Did I miss some hint for MNIST?

[gusty warren]

I thought it's easy too and thought there is only one thing you can count, count it in ways and for different sources and all came back try again....

[hardy barn]

pay attention to input_data

I really hope that people here for fair competition. Some people write with a proposal to make a deal, and it upsets me.

[terse moss]

Does the order of the input_data matter?

[naive umbra]

I didn’t check but probably

[naive umbra]

Unfortunately not

[gusty warren]

No joke. I respect the people that get 22+ flags. but the number of people that got 22+ surprised me

[twin wren]

the ranking always goes up together

[gaunt anchor]

I want to share my journey so far to the 23 flags .... its been a very very exhausting 19 days ... lot of failure and sometimes I spent days walking around challenges that I just needed a tiny small edit to solve ... I learned much ... so the 23 flags for me were hard ! ... and I am still hoping to get more 🙂

[minor falcon]

its not in the interest of people with high scores to provide answers to other people with high scores in most of the case. And there is too much at stake to risk a ban

[granite goblet]

well, now i count it in days

[hard perch]

Any advise?

Seems like I've got all the messages, still getting random strings

Seems like the blackhole people don't like me

[outer sundial]

Something trivial for you, might not be for others. Many people are stuck on MNIST for many days and that's not because it is hard, but they are not able to pin the correct solution.

[gaunt anchor]

Same as CIFAR most likely anyone pin the correct solution will get it ! .... its all about thinking ... and sleeping maybe !

I think most of the challenges here is to know what to do ! ... once you figure out, the coding part is not hard these days .. with GPT and other models around !

[ember relic]

I had the same thing, it turned out that i incorrectly typed in the coordinates

Recheck everything

[hollow coyote]

Yep, recheck everything lol. Little typos, fat fingers. They will get ya

[lost relic]

Hi, Granny 3 is extremely slow. I made a requests and it takes many seconds and sometimes even I get 504 errors.

[hexed lichen]

Is it a good idea to start working on this challenge with 12 days left?

[minor falcon]

most of us (in the 20-23 flag zone) got the flags in a week approx., so I'd say yes 12 days can be enough to go even beyond

[acoustic temple]

Finally got MNIST after 5 days and much sacrifice... 🥹

[wanton patrol]

nice, came back after a day's break to some memes

[acoustic temple]

granny3 endpoint is super super slow, zZzZ

is it intented? anybody has any idea ?

[minor falcon]

lot of people trying to find the weak pixel

[grave frigate]

Not me😌

I haven't even solved granny1😌

[severe pasture]

you can just use the granny1 endpoint for granny3 testing

[acoustic temple]

did you solve Granny3? I can't judge if this is a serious recommendation or just trolling 😄

[severe pasture]

it's not trolling, both endpoints return identical class predictions and since the prompt says that the objective is the same then a submission to granny1 that gets the flag can be submitted to granny3 (assuming the one-pixel rule is followed)

I'm assuming that doesn't count as a hint since it says it in the prompt lol

[glass bay]

well it's not a hint but it's giving an advantage away...

inb4 "why granny1 slow" messages

[severe pasture]

true, rip my <2s granny1 queries...

[minor falcon]

was also on granny1 😅

[hard perch]

For passphrase the same sentence gives different different scores. Anyone faced that?

[craggy beacon]

which one i'll test

[rocky jacinth]

Got MNIST.

[ember relic]

wow some characters are extremely stubborn for the ocr

[thorny ivy]

yes, if you make little changes like mayusc or more spaces, you get different scores, that's part of adversarial machine learning

[gaunt anchor]

I become more sensitive with 23 flags 😅 I think you are saying more than needed

[gusty warren]

like semi-colon and colon?

I get how it could be difficult for ocr, but come on we are in 2023, computer vision is supposed to surpass human....

[ember relic]

specifically i was talking about a different character

[gusty warren]

So one more character to worry about

[twin wren]

After 2 week

Still 17flag

[ember relic]

[gusty warren]

I should capitalize this CTF by competing in other comps while all the masters are trapped here

[thorny ivy]

Nice strategy

[granite goblet]

Oh god!!! Finally.... After 10 days, i have solved mnist

[sand veldt]

wondering how do you edit the timber_wolf image provided by host?? when I open it with PIL, even I don’t change any pixel, I can’t get the same bytes as the sample code does. The host sample code reads the the jpg as bytes directly. so do you edit the bytes directly instead of using PIL?

[severe pasture]

open it with PIL, apply whatever changes, save it to an io.BytesIO stream and convert the stream's value to base64

[grave frigate]

Has passphrase been solved?

[craggy beacon]

yeop

[grave frigate]

So only 2 solved them right?

I'm just trying to decide if I should even attempt or not😂😂😂

[abstract rose]

Inversion looks too unstable, yestderday AI wanted me to do playstation, today it's music. And I don't see any link with Ouija so far 😒

[hard perch]

Pickle seems so frustrating. I got into a new world after understanding the not the execution you're looking for part but still no luck. Really wondering whether the waving keyboard part have anything to do with it or not. Stuck being dangerous

[gaunt anchor]

Hello ... Hush server is down ? @olive ledge @limber flower ?

[gaunt anchor]

hush hush hush .... it seems the server is sleeping so huuuuush

[devout jasper]

it's definitely unbealivable I can't solve granny 1 and 2

I've almost tried every single library out there

[craggy beacon]

try something simpler like hush or passphrase

[thorny ivy]

Agree 👍

[ember relic]

LETS GO OCR

NOTHING CAN STOP ME

[gusty warren]

I don't think you need library for grannies

[devout jasper]

not even to perform attacks?

[gusty warren]

It's just identifying and changing pixels

you don't need a lot of lines of codes to do that

Of course you need library to read the img and modify the tensor. I assumed you were talking library specific for attacks, in that case, I don't think that's neccessary

[sturdy gorge]

There are some that can be used for granny 3

Tho they dont seem necessary for 1 and 2

[sand veldt]

Thanks. I did exactly the process you mentioned. Even without any change, I can't get the same bytes. I guess it's because host provided JPG has a compression level, so re-save will cause loss to every pixel. So I am curious how people can be sure the exact pixel value change they applied?

[severe pasture]

the images provided by the host should be png (are you sure you're using the latest notebook?) and you can save to the byte stream with format="PNG"

[sand veldt]

right, the name is png, but it's actually jpg, 'exiftool timber_wolf.png' can tell it's a jpg

[severe pasture]

Oh I see that yeah, regardless opening it with PIL, saving to a stream, etc and then submitting to granny3 doesn't trigger a "more than one pixel changed" error so the correct values are being received by the server

[sand veldt]

buffered = io.BytesIO()
img.save(buffered, format="PNG")
input_data = base64.b64encode(buffered.getvalue()).decode()

def query(input_data):
    response = requests.post('http://granny-pixel.advml.com/score', json={'data': input_data})
    return response.json()

query(input_data)
``` like this?  it returns {'message': 'I see more than one pixel difference.'}. 🥵

[ember relic]

Which challenges are still unsolved?

[severe pasture]

Weird, that exact code works for me locally - if you're using a Kaggle notebook I don't think you need to call decode()

[gusty warren]

Maybe your timber_wolf.png was overwiritten at some point

[sand veldt]

Thanks @severe pasture @gusty warren ! I tried on kaggle, the code above works as expected.

[hard perch]

Any advice on pickle? ☹

[devout jasper]

are you referring to granny 3 I think

[gusty warren]

They are more or less the same, with different rescritcion. granny 3 would be "It's just identifying and changing one pixel"

[ornate marsh]

be more dangerous 😏

[unborn sail]

As for me it is a very bad promt for task. Cause I've done more dangerous stuff, than required, before I got the Flag.

Is the flag for inversion meaningful? Or just a string. Even more, Is it in English?

[hard perch]

Banging my head with RCE

[gaunt anchor]

@olive ledge did you have the chance to check hush server ? its slow .

[unborn sail]

Is 457 still not triggered in inversion?

According to “wohooo I solved it in eternity” everyone is just bruteforcing?

[gaunt anchor]

Hush server is really slowwww .... 😦 😦 ... I am wasting time

[past brook]

read the hint and do some research

[gaunt anchor]

I am giving up on hush tonight… very slow … Going to sleep

[hard perch]

Well did a PhD on pickling. Still no luck

[waxen lynx]

Inversion slow and giving timeouts...

[sand solstice]

Tbh I should do a ranking of how “helpful” the hints were for each

Pickle is probably one of the worst offenders, but dont think it would be fair if they gave hints this late into the ctf

[jagged sluice]

pickle hint was the most helpful tf

just pick up keyboard and wave it

like a flag

[olive ledge]

It's a race to the bottom.

[ember relic]

thanks moo

[olive ledge]

https://tenor.com/view/wink-eye-wink-gif-12352742718877841322

[gaunt anchor]

hello ... can anyone confirm if hush is slow ? or it's something from my side :/

[gaunt anchor]

😫 😫 it takes around 40 sec per request

[craggy beacon]

yep

Hush In a world full of clamor, where noise does abound, There's a challenge that waits, where silence is crowned. Not for those who roar, or shout out so grand, But for those who speak softly, and softly they stand.
In whispers and murmurs, the secrets lie deep, While the boisterous might struggle, the muted shall leap. Loudness won't aid, in this quest you partake, It's the soft-spoken riddles that keep you awake.

So tread without noise, make not even a peep, For it's not the loud mountains, but the quiet valleys so deep.

[modern anchor]

Hush is unsolved I guess

[gaunt anchor]

No Hints, and even if someone can speak loud his/her thoughts ... it may lead you to wrong directions .... its just what we have on the notebook

@olive ledge @limber flower I know I ask much about this 😅 please wake up hush server 😄

[grave frigate]

😂😂😂

I meant

It's so🥲
I meant to say were there any hints in the discord chat that I could hunt for

I've already lost 2 weeks to pickle😂😂😂

[grave frigate]

This worries me
Like that day a guy was bashed for it

[gaunt anchor]

I don't think there is anything here on discord that you can't find out your self while playing around with hush

but the server is really slow ... makes it hard to even experiment anything :/

[grave frigate]

For discussion about gradient descent

😂😂

[fluid robin]

Does pickle needs bruteforcing?

[unborn sail]

actually no, but approaches should be tested

[sand solstice]

in my experience, all challenges can be solved via bruteforcing, but in practice intuition played a role in most of my solutions

[minor falcon]

mmh i don't agree, there is a bunch of challenges that cannot be brutforce

[final path]

given limited competition (and probably universe) lifetime - yes 🙂

[modern anchor]

For anyone who solved inversion did u brute force values of 4,5,7 index characters?

[queen garden]

Why would anyone who solved inversion answer your question ? Giving hints is against the rules.

[native jasper]

Hello

[outer sundial]

https://tenor.com/view/sad-cat-gif-26527456

[modern anchor]

That's fair. Was just frustrated. 😭

[acoustic temple]

hmm, so it seems I'm quit on wrong direction, since I even can't get any high prediction on index 2, 4, 5, 7, or maybe not

[lyric token]

are you sure that you right predict 6?)

[modern anchor]

Not 100% sure but surely more than 4,5,7

[orchid oriole]

I hate MNIST, it's so guessy, in every other CTF I wouldn't waste my time with it.

[lyric token]

Yeah, I hate it too even if I solved it

[minor falcon]

i found mnist more than ok

[final path]

me too, but I think it depends on how much time you spend on it. i still hate CIFAR 😄

[minor falcon]

yeah cifar is another league :p

for multiple reasons actually that i cannot mention here unfortunately even if I really want to aha

[final path]

next year: simple counting challenge for Imagenet

or The Pile

[minor falcon]

i suggest a counting challenge on the weights of a pretrained mobilenetv2

[orchid oriole]

not if you forget for a second how easy it was to solve for you personally and think about the path towards the solution.

Sure, it's super easy if you have the right intuition. But there's nothing guiding you there.
There is no process of gathering bits of information and putting them together to solve a riddle.
You don't get feedback from the challenge so you cannot reason about how right/wrong your approach is.
There is no (meaningful) hint to start with (the only 'hint' didn't help me at all and other people complained on here that they found it misleading.). The right solution has no inherent meaningfulness, which could help in finding it by looking at the dataset itself.
The only question is, can you guess the arbitrary metric the challenge author thought of, or not, then try again. And even if you finally manage to guess the right thing, you haven't learned a thing. That's just lazy design.

I really like the comp as a whole, so no disrespect towards the authors at all, just thinking about CTF challenge design.

[minor falcon]

i think there is more than enough information for mnist and proof is: all the people who tryied to compete successfully got the right answer. With a good methodology on that one, the search space is not that big

a challenge like inversion might give a lot of info, but if you dont see the trick, you are doomed to fail as much as for mnist

[orchid oriole]

that it's easy to guess is no proof that there is enough information^^

[minor falcon]

and i can speak about it as i'm hurting a solid wall on inversion while it seems like all my mates from the prev 22 clubs managed to solve it :p

[orchid oriole]

the only argument I see here that I can accept is, if guessing the right metric until the outcome matches the expectations is data science, then this is a good data science challenge, yes 😄

[valid cobalt]

I stuck at guess who's back since 3 days and people say an easy one 🥹

[outer sundial]

https://tenor.com/view/again-guess-whos-back-again-eminem-rapper-hip-hop-gif-6224057

[minor falcon]

what I mean is that a challenge can give you a lot of information, but if you don't see the right trick, you might have the best methodology and best models, you'll never get the right answer.
Mnist is a bit like this, but unlike other challenge where you need to 1. see the trick 2. build a model around, for mnist, you just need to "see the thing", so its somehow even more simple than other challenges that gives you information :p

[past brook]

the search space is big, there are infinite ways you can count in depending on how you interpret the hint

for me, I interpereted the hint "incorrectly" which made me waste days of work for nothing

[orchid oriole]

nah, if the connection between initial information and solution is obvious in hindsight, it's good information. even if It is super cryptic at first and you need 50 hours to understand it or to get the right idea, that's more a 'you'-problem then :p
And in the aftermath, you probably learned some technique / way to look at data / whatever (at least a little bit about how the author makes up hints), assuming that the initial bit of info is vaguely pointing at some inherent characteristic (otherwise it's hard to provide any hint at all).

On the other hand, you've got NOTHING, your only tool is guessing in the dark.
Even if it's easier and you solve it in a fraction of time, that's not fun (to me at least, and most people in the security CTF scene agree I'd dare to say)

[final path]

I think some feedback like number of correct entries in array would make these challenges way less frustrating, while keeping them almost as complex

[minor falcon]

More easy to brutforce also unfortunately

[wind ether]

I'm still with you 🥲 . Honestly at this point I refuse to look at inversion, I figure unless I solve 1-2 others there's no point really since I'm already out of gold range

[final path]

brute forcing 500 random numbers is not possible even with this kind of feedback 🙂 and you can always make 10sec response time to avoid it

[craggy beacon]

how much time did u spend on cifar and have you learnt somethingusefull from that?

[minor falcon]

learning something is not necessarly the sign of a well made challenge :p
I did not learn a lot of stuff doing "guess who is back", but i had a lot of fun doing it :p

[acoustic temple]

tbh, I did learn (or at least I did practice) something while solving MNIST

[final path]

im now guru in manipulating multi-dimensional numpy arrays 😄

i think i spend ~5-10 hours actually trying something and coding

[glass bay]

it'd be great if just the hint was a bit more transparent, like saying "here's a one row out 256/100 that are needed"

[minor falcon]

that I agree

[craggy beacon]

because resources you spent is proper for the reward you get

other, more complex tasks can be more rewarding because you learn something on the way, so more rewards for more resources spent

[minor falcon]

not as detailed as your proposition, but more meaningfull (people who solved mnist know what i talk about)

[glass bay]

like honestly that would be 100% enough for cifar to become solvable because you enter a loop of "what could have a form of 'x,y,z,w'" and at least have some gradients in your brain from how logical that seems

[minor falcon]

well i think for cifar this is more or less known for people that worked on it the "form x,y,z,w"

what is less is a clue that could definitly indicate you are going toward the right direction

[craggy beacon]

hints are cryptic and lead to vast search space, but when you find solution you understand that hint was actually ok

[glass bay]

oh, i think i have a way to make cifar solveable, explain the solution to GPT-1 like it's a 5-year-old and ask questions from it

[craggy beacon]

so less reasons to hate it

[glass bay]

search space can be greatly reduced with my example, allowing you to go 'aha, the sequence "2, 4, 8, 16" may mean that or that or that and i have at least something to search for'

[minor falcon]

I'm 99% confident about the logic behind the input sample for cifar
i am 75% confident about the data shape
but yet, there is one piece missing

i think it would be reduced too much with your example

i think its like mnist, if you give slighty more you make the problem trivial to solve

[glass bay]

imo what would be fun without trivializing is to make an answer something like 'find labels of digits in mnist that have a hole in them(4 has a hole and 5 does not)' and form a hint around that. you'll be actually interacting with the data provided and will be forced to do something with the data that is more complex that counting something somewhere

further more, the concept of 'counting' is so vague, that its impossible to be made interesting and complicated

imo*

either it's one of the first 5 guesses you come up with, or one of the last 4

[ember relic]

personally i found bruteforce to be very useful for some problems

it helped guide me in the right direction

combined with a few educated guesses it turned out to be very useful

[hard perch]

Is the waves keyboard clue really helpful for pickle?
Wanna bang my head with keyboard 🥲

[sand veldt]

what's the sota for granny3? no flag at 0.000831, I feel empty

[minor falcon]

i think that 0.000831 is the sota :p

at least its the highest value I see here, we are many people "stuck" at 0.0006927

[sand veldt]

oh, thanks, I start to doubt if maximize score is the right direction...

[minor falcon]

apparently you already managed to figure out something that many of us don't have, so you are probably in the right track

[sand veldt]

hopefully, but I feel the seach space kind of exhausted, so maybe optimization goal is wrong

[grave frigate]

Has cifar been solved?

[final path]

maybe, but not by me (hopefully .. yet)

[jagged sluice]

yes

pick up your keyboard and wave it around

[minor falcon]

how do you do when you have a laptop?

[jagged sluice]

well no flag for you then

unless your laptop is modular

then unscrew keyboard and wave

[acoustic temple]

so does it mean it is not one pixel, but doing in a way to pretend it was 1 🤨

[random minnow]

after reading about 50 to 60 papers, there is more than what you all think. ask google, search widely

[gaunt anchor]

I think hush hates me ! ..... slow slow slow .... more than 24 hours and I didn't progress 2% of what I wanted to do :/

[wanton patrol]

going against the grain and will say that the more I thought about it, the more I liked the MNIST hint 🙂

[modern anchor]

Still didn't get it even after solving🤔

[gaunt anchor]

Congrats @craggy beacon 😉

[severe pasture]

@craggy beacon which one did you solve (or 🤐 )?

[craggy beacon]

not granny3

[wind ether]

Passphrase?

[final path]

🕵️ from the chat messages i think the last one was inversion & cifar solved before 🙂

[strong pilot]

Semantle - Level 2 --> done

[violet trellis]

Hello guys
Just want to know, does the tensorflow/pytorch version affect the model predictions/weights somehow? I mean big differences, not 0.00001 ones.
I just have some hard time replicating granny model.

[craggy beacon]

maybe
though I hate inversion and it is a bit easier for native speakers

[severe pasture]

big differences are expected

^

[hollow coyote]

Just when I thought I had MNIST, "try again"

[gusty warren]

I saw the same question in another compettion. My answer is no, it is probably something that you didn't match (default parameters, processing pipeline)

[severe pasture]

nah the organizers specifically said to expect large differences

[gusty warren]

This is the right answer

[sand veldt]

the problem itself doesn't say it must be one pixel. and the server could return "'I see more than one pixel difference.'" and "flag: {}" at the same time🧐

after kind of exhausted the 1-pixel search space, I start to consider "I see more than one pixel difference" is just a troll

[severe pasture]

even with regular preproc the results will be different

[gusty warren]

The many people that matched the model wouldn't agree, unless you are talking about granny 2

[severe pasture]

my point is that there is additional preprocessing happening beyond the standard resizing arrays and scaling values that needs to be contended with - the question earlier was asking if, when using the model normally, they should expect such large differences, which they should

[sand veldt]

wow another 24!

[lost relic]

That should not have sense in my opinion

[sand veldt]

curious why?

[lost relic]

If I remember well the "'I see more than one pixel difference.'" has a 400 status code, which means that the client made an error

I don't see the point of trolling us with that, it is not the spirit of the challenge 🙂

[gusty warren]

" there is additional preprocessing happening beyond the standard resizing arrays and scaling values that needs to be contended with", that's a strange statement from my perspective. But I am afraid I would be giving away too much information if we continue discussing this.

[glass bay]

Host said that the intended solution is done without model copying

[devout jasper]

what does it mean model copying?

[glass bay]

As in use the return values from the api to try to approximate whatever they have going on to have stuff that api does not give

[gusty warren]

That's true. But at the same time, the host also specified the exact model used in the prompt (presumbly to lower the difficulty) and it would be 1000x time faster if you can match the model

[glass bay]

...if the matching won't take 10000x longer in total and will be precise enough

[gusty warren]

True, depends on one's search method, it can be faster either way.

I intended to solve it as black-box, but the api response time kind of forced me to change strategy

[minor falcon]

Its not so bad the api response time atm. With the current response time and the right method, the task is achievable in a descent time

[thorn quarry]

@olive ledge does inversion contain only letters or it includes special characters and numbers too??

[random minnow]

if you are still stuck at pass-phrase

[craggy beacon]

did u find something usefull on CIFAR or Hush in the chat?

[viral bough]

i just kept entering manual prompts and got WTF 6. I have no idea what kind of prompt injection or anything 😄

I wonder if others had the same experience? Was it the same of the other WTFs as well?

I haven't gotten 3, 4, 5

[craggy beacon]

wtf are quite random

[rocky jacinth]

Rather enjoying chatting to WTF5. No clue how to solve.

[viral bough]

Same here. I got WTF4 randomly now as well. But, can't get past WTF3 "I know you are, but what am I?" and also WTF5

[neat light]

I am getting {'s': 0.0} on cluster 1 and no flag...I guess the target was {'s': 1.0} then?

[viral bough]

I wonder if there are some systematic prompt injections we can use for WTF

[random minnow]

i suppose after the competition, the server will be stop. any chance that the server setup/file (e.g. score.py ,wsgi.py, docker setup etc) will be made public and available after the competition (for post usbmission, practice, etc)?

[sand elbow]

Last year they left the server up for a few days after the competition

[random minnow]

i would suggest to keep the server for at least few weeks for those who want to show more experiments results in solution writeup, etc

[sand veldt]

wondering 22-hackers, what are you stuck on?

[fallow valve]

Inversion

[gaunt anchor]

My experiments on Hush failed .... big time ! ... currently I am clueless ! 😑

[modern anchor]

I felt like a real hacker after writing the prompt for witf3 which got me the flag.

[random minnow]

"wondering 22-hackers, what are you stuck on?"
actually they are not stuck. it just that the remaning like granny3, inversion, passphrase takes more time to crack as these involves building some model to solve them

the search space are also large

"My experiments on Hush failed" ... hush server is too slow. and many of use are not so failiar with sound processing. amybe other probalem is a better bet

it takes time to find "more information", e.g. from experiment observation, etc

[unique hedge]

got 20+ sentence about passphrase......

[acoustic temple]

passphrase is good

now you have 21+

[unique hedge]

lol, but 20+ sentences for "where everything is equivalent"...QAQ

[acoustic temple]

in case you missed this, @unique hedge

[unique hedge]

Oh, thank you. Now passphrase become a secret sauce...

[empty bane]

Did you solve it?

[gaunt anchor]

When I though I had CIFAR ..... and the output ... try again .... I should accept my current position and accept silver

[devout jasper]

same for me with granny and pixelated...I feel dumb every time I read this chat

[past brook]

its because people assume everyone already knows everything

[final path]

try passphrase, at least we know it's solvable 🙂
but I'm starting to have a headache reading this promt for 1000th time...

[sand solstice]

proof?

[final path]

this

and some messages above

[gaunt anchor]

I am .... for 10 days so far and still can't get it right ... but atleast it server is faster than hush hhhhh

[glass bay]

not really a proof since for organizers everything is solveable, just that passphrase was solved by someone in discord

[craggy beacon]

Do u think cifar is not solved?

[gaunt anchor]

New plan, stop all colab notebooks ... once I reach the house sleeeeeeep then start from scratch with passphrase and CIFAR ....

[final path]

we don't have strong proofs from non-ogrs about CIFAR 🙂

[gaunt anchor]

I think CIFAR is solvable once you see what to do .... it will not most likely take time if you have the right direction ... and most likely was solved

I say "most likely" much ... AI AI AI

[final path]

actually, it's the same for passphrase. when you know, you know

[gaunt anchor]

Yes same for passphrase .... there still 10 days so maybe we will see more 24 and 25 I guess ....

[sand solstice]

assumed this was a joke

I think the concensus is cifar has been solved by a few, not sure about the other 3 "big ones"

[gaunt anchor]

@gusty warren has solved passphrase , he mention it earlier

[pallid trench]

any hint for semantle 2

[acoustic temple]

it is Semantle basically, but with 5 words 🙂

[random minnow]

"any hint for semantle 2" : check google/youtube on how people solve semantle

if you can't solve it yourself, search for more information

[unborn sail]

is inversion server down? every 100th i get timeout! Even with time.sleep

[acoustic temple]

which one, ouija board endpoint? or flag endpoint?

[unborn sail]

Flag endpoint is soo slow

[sand veldt]

thanks god chatgpt, I finally understand the story of pass phrase. Instead of asking gpt to translate, I asked it to do reading comprehension like toefl test. now it's crystal clear

[median mauve]

@exotic flame why did you share the code for granny?

[olive ledge]

@exotic flame thanks for removing. Keep on CTF’ing 🙂

Awesome notebook too. Looking forward to reading through all of them post competition.

[devout jasper]

stuck there for almost 10 days...missed my opportunity 😦

while others, apparently, got it

[unborn sail]

@olive ledge can you please tell the situation on inversion Flags' endpoint performance? Looks like it is dying rn!

[olive ledge]

Checking…

[devout jasper]

anyways, it remains a mistery to me why Kaggle can't disable the notebook section for specific competitions

[minor falcon]

i saw the notebook, its intersting, but if you documented yourself a bit, you probably bumped to the same kind of stuff already 🙂

[devout jasper]

I've documented myself a lot, for sure, obviously without success. The fact that others saw that notebook and probably solved Granny frustrates me a little

[minor falcon]

i just saw it very briefly, but to me it was more focusing on grany3, and definitly not usefull to solve grany1/2 (not even 3 i would assume - but again, i passed through it very quickly)

[outer sundial]

It contained solution for granny 3? 🤔

[devout jasper]

I hope not even the solution for 1 and 2, otherwise I would expect some actions from the hosts

[exotic flame]

It doesn't contain no solution, just an implementation of an evolution strategy, not enough to solve granny 3 guys

[devout jasper]

thanks for clarifying!

[exotic flame]

I think that we need another kind of effort to solve it, maybe more subtle, I didn't solve of course, I am not crazy 😄

[minor falcon]

and for the people struggling with granny1/2: read research papers. A lot of information is available out there explaining how to solve that type of tasks, and the algorithms are not that difficult to code by yourself

[past brook]

My issue with granny is I know what to do in theory, but Im too lazy to match the model

[random minnow]

"but Im too lazy to match the model" there are both black box and white box method

[sand veldt]

imo, for granny 3, the strategy doesn't matter, brutal force is enough, search space is actually way smaller than one expect. the key is what to optimize for, same as other challenges.

[past brook]

I solved all of them with the same query 😄

[unborn sail]

Inversion totally down!

stopping my 10 notebooks😂

[sturdy gorge]

trying out granny, im not being able to submit the query as it is shown in the starter notebook, its just gives Object of type bytes is not JSON serializable

[light mulch]

input_data = base64.b64encode(img_bytes).decode()

does that work?

[sturdy gorge]

That works perfectly thanks!

It was missing the decode

[sturdy gorge]

is it really mobilenetv2 for the granny challanges?

[gaunt anchor]

0.0007727496 granny3 hmmmm

and the code break :/

[acoustic temple]

I saw somebody posted something like 0.0008...

[gaunt anchor]

ya I saw it ....

let see if I can break it

[acoustic temple]

here @gaunt anchor

[lost relic]

Just by modifying one pixel?

[gaunt anchor]

yes

I need to check ... i smelll a bug

[ember relic]

maybe you got the flag and it crashed?

suffering from success

[gaunt anchor]

no no ... I get a wolf :/

I can't re generate .... I think its an error ... forget 😦

I think we can do it ... I start to like granny3 .... hate hush

[sturdy gorge]

feel like im going in circles with granny1, maximum i got was 0.0014

[ember relic]

same, its probably the overthinking as usual

[sturdy gorge]

got the moled to classify as a granny smith, but now the second probability is a tennis ball, 0.01 ahead of the wolf

hum, getting classified correctly, but still no flag, interesting

[gusty warren]

Dude, that's what I got when I used Microsoft paint to make a submission....

[sturdy gorge]

Haha, i tried something different, but maybe the problem is similar

[honest zenith]

@minor falcon what papers do you recommend to start from?

[minor falcon]

its part of the game to find out by yourself.
But i'd say:

1. start with vulgarisation articles to build yourself a high level culture about the subject
2. when you understand well what needs to be done and the different families of methods, you'll have to look for papers that propose implementations under certains constrains that you would have identify. Check for multiple methods, and make a trade off between stuff very powerfull but hard to implements, and stuff easy to implement but not super effective.
3. Lot of people have probably already coded stuff that could be usefull, so i'd also advice you to see if there is not a public repo that implement (partially or completly) what you need, once you have figure it out.
4. Try it out, and monitore what you are doing to see if you are potentially heading toward the right direction

[sturdy gorge]

I have curated a set of cool resources, but without being able to replicate the model localy, they dont seem much helpful

[violet trellis]

another 24

and i still can't pass 18 for 2 weeks :/

[minor falcon]

maybe there is other directions to look at. Orgs already said that "you don't need to do this. In our solution for testing, we don't use a surrogate model". It has been posted a few times here.

[hard perch]

Any advice on pickle please? Except waving to keyboard🥲 been less dangerous and more dangerous. Seems like nothing works

[sturdy gorge]

yeah, i ahve been trying another approach but even so, it doesnt go above 29% and thus the flag is no popping sadly

if by any chance the model can be replicated localy tho, there are several out-of-the-box algos for granny 3

[past brook]

20!

late comeback 👀

[random minnow]

"there are several out-of-the-box algos for granny 3"
are you sure? if you search the literature for sparse attack for imagenet (not the small cifar), SOTA needs more that 1 pixel

[sturdy gorge]

Not claiming they are the current sota, but I did find several one-pixel resources
Looking at the papers it seems it should work, tho in practice that can be a completly different story haha

[lost relic]

@olive ledge @limber flower I believe there is some problem with hush security

I'm sending a bunch of audios from my pc and I'm receiving 403 ret codes very frequently, despite making very little number of calls

Or maybe the APIs share the same security and I'm being blocked for my calls to granny XD

[devout jasper]

what are all these numbers you'all are talking about for granny 1 and 3

[gaunt anchor]

I had nightmare with hush response time .... and I gave up

[minor falcon]

probability of the targeted class

[gaunt anchor]

I ran out of ideas for CIFAR, passphrase , granny3 and the so many days wasted "hush" .... I feel my head is empty now !

[sturdy gorge]

dam, would really like to get granny 1 and 2 before the end of the comp, but cant get the probability high enough for it to trigger 😔

[icy notch]

in granny 1 I have experiments where:

1. I have received granny as first prediction and wolf as second
2. I have received wolf as first prediction and granny as second
3. granny have more than 0.7
   but still no flag...

[sturdy gorge]

hum, that even more concerning haha
then for sure im following a wrong approach

[minor falcon]

As it has been said here in the past, posting an image of an apple won't give you the flag, even if you get a ridiculously high probability.

[icy notch]

in my experiments I changed the wolf picture using some black box method

[gaunt anchor]

A probability such as 0.99999989x If I may say 🤣

[sturdy gorge]

ah, i saw those posts before, and was not providing the image of an apple

[minor falcon]

this is just food for thoughts, after its you to figure out whats happening, what works, and what doesnt

[sturdy gorge]

yeah that fair
currently im just lost in what else to try

[random minnow]

i find an interesting attck for future competition: fool an object tracker

i also think of the story: you are being tracked and targeted by enemy drone/missile. you have 5 second to act, think of a way to escape the tracking missile

[acoustic temple]

I don't think the challenge would be as clear as you wrote, probably it gonna be something like this (GPT generated):

In shadows deep, a tale is spun,
Of danger near and foes well-spun.
A hunter's eye, a target found,
In moments ticking, danger's bound.

Five beats, a breath, a fleeting chance,
To dance with fate, to swerve and prance.
Escape the eyes that pierce the night,
Before the missile takes its flight.

In whispers low, a plan conceived,
The path to safety, yet deceived.
What will you do, how will you flee,
From danger's grip, will you break free?

[tribal plank]

Try again!

[random minnow]

this is also my favourite attack in video: dirty road can crash car

[light mulch]

I'm probably missing something, but is Cluster 3 on purpose misleading? Either I get "No credit, please top up" or i change something i think i can read from the data and get {'message': 'Invalid input.'}.

[sturdy gorge]

It should not be misleading

seems like you got something wrong

[rocky jacinth]

The input structure in the given starter notebook is correct, those are the three fields you need to fill. Any indication otherwise is misleading.

[light mulch]

Thank you! Thats what I figured from {'message': 'Invalid input.'}. 😄 The data just is so convincing to do something else.

[rocky jacinth]

Indeed.

[hard perch]

This thread might help

Been there. Keep trying. You’ll eventually get it

[light mulch]

Thanks! I knew about the possibility of reading the characters wrong, and I just now found that indeed my issue was related to 0 and O . Still have no flag but at least i have credit. 🙂

[hard perch]

Best of luck! May the extraterrestrial people be kind to you

[thorny ivy]

Cygnus X-1 is proud of you

[hard perch]

Cause you are not that dangerous ☠️

[modern anchor]

I am very dangerous but still can't get past pickle🥲

[random minnow]

note that the objective of this competition is "Collect flags by evading, poisoning, stealing, and fooling AI/ML"

[glass bay]

... except count mnist/cifat

[random minnow]

e.g. no point sending an apple image as it didn't fool anyone.

hence score, etc is not the only factor.

what we really one is to attack the AI/ML, i.e. we want them to make "silly" mistake

[abstract rose]

Inversion is resisting, I'm wondering if only native english guys could solve it.

[gaunt anchor]

I am not native english speaker ...

[ember relic]

also solved it even though im not a native english speaker

[gaunt anchor]

I am going to hush again ... hope this time the one logical thing I can find become a flag ! ... which didn't earlier but maybe I was doing something wrong

[thorn quarry]

I was trying something with the hush, length of output for hush changed from 8 to 2, but no idea why did it happen

[hard perch]

For granny 1 and 2, do we need to submit the provided timber wolf image with certain modifications or we can choose to submit other images too?

[sand veldt]

omg, stuck on passphrase too long! There is only 1 spot left in prize zone, meaningless to continue if that spot is filled

[gaunt anchor]

I wouldn't look at it that way .... there is still a week and who knows .... maybe we will end with 25 or 26 .... noway 27 hhhhh

[sand veldt]

I wish to be as optimistic as you if I got passphrase

[modern anchor]

Same happened with me as well

[sand veldt]

Hush length can be any number from 2 to 8

[modern anchor]

Have you solved it?

[sand veldt]

no, I just have some sense of the meaning of output

[minor falcon]

can be more than 8

[sand veldt]

yes, typo, any number from 2-12 in my trials

do you find anything larger than 12?

[thorn quarry]

No,any idea what those numbers mean??

[sand veldt]

can't say. you can brutal force like inversion

[acoustic temple]

I did not solved inversion, but I can't see relation between Inversion and being naitive English speaker

[abstract rose]

Oh, some people here solved Inversion by brute force the flag endpoint?

[final path]

2 people who solved inversion first were not english-native 🙂

[final path]

i think there will be 25 (we know that passphrase is solved by at least 1 person and CIFAR probably by a few [but i refuse to believe])

[minor falcon]

I tried more than a million combination without success

[acoustic temple]

I think it doesn't make sense, one need to tight search space, otherwise it is not fit in timeline of competition to brute force it 🙂

[abstract rose]

We even don't know the length of static flag, it could be 3, 8 or more

[craggy beacon]

agree 25 or 26

[ember relic]

passprahse has been solved?

why are you so sure

[final path]

^

[ember relic]

damn i missed it

thanks

[craggy beacon]

i think only granny3 hasn't been solved yet

[acoustic temple]

hush?

[ember relic]

granny3 this years crop2 hahah

[minor falcon]

Yeah i tried different strategy, i even thought i figured out the way to get the missing info, but apparently not 🥲

[thorn quarry]

Does test_noise has any help in solving the hush or its there just to show the input format?

[random minnow]

should be 12 (could be words or charcter?). doesn't make senses to ouput N numbers if the target is M, from design point of view

[final path]

it's just an example, according to orgs

[unborn sail]

оh nice than, I spent 2-3 days denoising it(

[wanton patrol]

same, was kinda pissed 😄

but it literally says test_noise

so cannot blame anyone but myself

[unborn sail]

stupidity is my main enemy

[ember relic]

i really hope its not words otherwise we've got semantle 3 on our hands

[thorny ivy]

Lol but I've found words denoising it 🥲, so I'll try some other things

[sand veldt]

"[but i refuse to believe]"

[rocky jacinth]

Semantle 2 - done.

[sand veldt]

one likely hypothesis: 24 got passphrase, 23 got inversion, R.I.P. for folks trapped in 22hell including myself

[minor falcon]

wait to see the solution of inversion, thats where the pain will be the hardest

[ember relic]

i like to think 24s got cifar