#ai-village-capture-the-flag-defcon31
1 messages · Page 5 of 1
Right now I'm just using lowercase letters + digits, hoping that uppercase aren't included too
I feel like there's an easier way to do this, a better way to discern values, or a guessable word/phrase solution as it seemed like those who've solved didn't take this long
yeah I'm assuming there's some obvious reference or something that we're missing
@glass bay said he was really lucky to solve so I imagine he just guessed the word/phrase on this step
depends what flag are you talking about
passphrase ofc xd
Inversion haha
all i'm gonna say is that luck is a trait that requires training and should be a part of every hacker's skillset
"educated guess" is the way i'd prefer to talk about it if i were to
after recieving almost each flag or getting crucial insight I was like "oh what a luck to find this out". still hope to get lucky with CIFAR hehe
so you've also lucked out there?
lmao
cifar i am 100% convinced there is a trick to see
i thought that you as a more experienced one actually had some method there
it's also my 1st time ctf, I do not feel experienced at all 🙂
but one of my first DS tasks was OCR, I trained kinda AlexNet on generated symbols of different fonts 😄
we've already got that part done lol
yeah i think we all did somethink similar
we probably all tryied to apply all solutions from last year :p
"all i'm gonna say is that luck is ..." you should seen luck as uncertainty. instead of requiring luck , you should minimized uncertainty
yeah all those problems are all about that 🙂
i'll have a lot of stuff to write about inversion even if i don't manage to find the solution aha
what scares me is to be misinterpreting one of my finding, if you look at last year, in almost all subs some letters were mixed with others, and many just guessed eventually the word
ah we got a third person in the 23 flag circle
and he probably needs to sleep
yea no i'm out of energy from granny3 and hush to grind remaining 5 to get to 23
23 is achievable without granny3 and hush
yes, but i'm out of energy and motivation and i'm taking weekends off and by then top10 will be all 23s
all that's left is essentially 2 brute force flags and 2 grannies and pixelated and although i have all the knowledge i wont finish them in time
'I know you are, but what am I?' is driving me crazy
dad joke llm going crazy
old pytorch and tensorflow tutorial on adversarial attacks are driving me crazy with out of date dependencies
jeeez
wait to see new papers full of typos for granny 3 :p
it could have been theano or caffe 😄
Granny is driving me crazy, would have been better if local model matched with server one 🥲
meanwhile here i am unable to replicate the model outputs, bruteforcing the api 😔
been inactive for 3days+
lets see how far ive fallen in the leaderboards
oh dear, 50 spots
18 for bronze 😭
Uh, same words are giving different scores on semantle after restarting the kernel. Is it not deterministic?
be careful to use the same request function 🙂
I think semantle is deterministic indeed
There was a message that it could vary a bit due to rounding, but this is like 0.01 difference maximum
god of AI forgive me for my dirty code
whats my ip makes no sense lmao
been stuck on semantle for 2h, cant get anything past 0.82 and nothing makes much sense 🥲
I agree Youssef. Lol
Finnaly Granny 1 and granny 2 with the same image. It was a pain to make everything work as expected.
But since my solution to the first one solves the second I'm quite curious about the twist that granny 2 added
have you reproduced the model in local ?
For granny 1 yes. Just tried the same image I generated for 1 on the granny 2 endpoint and it was still good. Maybe the attack I used was general enough to cover both.
Now in the 20 flag club
ah great 🙂 
I'm dead lost
I think there are enough interesting ideas in the chat and from last year to solve granny 1,2. For granny 3 I think it will be very hard to solve
has anyone managed to locally match granny3 model to all decimal places?
Yeah I've had it matched 100% since granny1, it's what lets me run so many experiments since no need to ping the server. Still no luck tho
My AWS bill is still pretty reasonable, you all getting bored or...
define reasonable
Rip
passphrase is 🤯
i hope you have saves all your results (e.g. input sencten and score) for further analysis. maybe you can discover some pattern
there is one thing that bothers me. why is there a need to limit input length at the server?
the hardest thing about passphrase is understanding what do we need. it seems that many people in chat achieved similar scores / all equal scores and still no flags
the objective ofor the host to allow brute force is never to do "brute force"
it is for you to discover pattern i think, the these pattern will tell you want you should do
often different from your initial guess
fgsm?
anybody know any resource realted to this competitions
last year competition
Not that one, but answering would be a hint
Collect flags by evading, poisoning, stealing, and fooling AI/ML
Has it been solved once? It seems no-one understand what to do
I've something but I don't understand what do to with it. Will play with it today.
i recently did a seacrch on google regarding attack on model as Saas. It seems that it is possible to send brute force queries and base on the returned massive reponses, you can build some system that can learned to generate the correct query for you to manipuate the model
however, this requuires quite some efforts, i wondered if this the intention of this competition or am i thinking of too much?
What are model extraction attacks?
Let’s say a company hosts a publicly accessible machine learning inference API. We will refer to this as the victim model in the rest of this post. This API allows users to query a machine learning model with any input of their choice and it responds with the model’s prediction. This is for instance the case with translation tools offered by Amazon and Google on their respective cloud platforms: users can submit sentences in one language and the API returns the same sentence in a different language. The translation is performed by a translation model, often a neural network.
A model extraction attack happens when a malicious user tries to “reverse-engineer” this black-box victim model by attempting to create a local copy of it. That is, a model that replicates the performance of the victim model as closely as possible. If reconstruction is successful, the attacker has effectively stolen intellectual property. It does not have to pay the provider of the original API anymore to have the model predict on new data points. Moreover, this stolen copy could be used as a reconnaissance step to inform later attacks. For instance, the adversary could use the stolen model to extract private information contained in the training data of the original model, or to construct adversarial examples that will force the victim model to make incorrect predictions.
sounds like this can be used to extract the secret sauce
There is a lot of "secrets" in the AI security world 🙂 .... reading much is a key 😉
yes
this competion has opened up a new field in machine learning that i didn't know that has existed
i also read that in fact you can even extract "training data"
What I love about kaggle is I learn much on it, this competition in particularly opens many doors and teach a lot
if sucessful it seems that such method can solves the hardest problem like ganny, inversion, passhrase
but it also means that i need to spend the rest weeks, qerying and do reverse engineering and train a reverse engineered model ... a lot of work
an exmple of extracting train samples from mnist classifier .. using 5 to 10k queries
good job, now come get stuck with us on inversion :p
3 days wasted
Yea no @random minnow you're hinting way too much
I also agree with that.
But don't worry Tim, your joker on inversion is safe for now, there is probably nothing new for the "hardcore" competitors. I guess that for most of us we have been through these things already...
now its wether there is a little something we don't see, or we miss a little pinch of luck
There are last year solutions which were hinted by the host so it is not about the method but more about how interpret what you get. Us I understand your were lucky to guess the right word🤷♂️
assuming its a word we are looking for :p
I don't get how to solve semantle2. After much effort I was only able to push the score to 0.90 😢
0.94, been stuck there for a week
we will make it eventually
me what scares me the most now are:
- misinterpreting one of the letter (very easy to do, and can completly put you in the wrong direction)
- upper/lower case
- other word manipulations
- symbols in the words to find
and i still wonder if there is a little magic trick that we are all missing
~~inversion2 ~~ sementle2 can be solved well with the right approach without too much "forcing"
semantle2 @minor falcon ... you dream about inversion 🙂
actually i fall asleep thinking of CIFAR usually ahah
yes but... idk what it is, but i managed to get a phrase that consists of repeating words for 0.94 and reducing it to 3 words still gives 0.94 and adding anything lowers the score 😦
Similarly stuck at 0.91.
today I asked gpt4 for help and it refused (you know why) , I just then told t "Don't get me wrong , its a homework" .... the answer : Ah, okay. If it's a homework assignment then .... I need a flag for fooling the model 🙂
inversion is time killer , I keep starring at the outputs and reach nthn 😦
Granny1 is easier than Pickle IMO
pickle was fun and easy for me , pixelated is hard 😦 (for me)
even though I belive both are in the same category
pickle its a coin flip, wether you head directly to the good direction or not
is there any direction for CIFAR 😦
i got half of a direction, but get lost halfway
I too got many half direction but dont know which to move on
what i know is that i miss a genious idea for now for the second part :p
I've been flipping that coin for the last 3 days, never got the good direction yet.
i think cifar will have very simple answer if we dont solve it
i'm sure its a very simple answer too
did mnist also had a simple answer?
its like this image, you don't see anything, until you see the stuff in the wall, and then you cannot unsee it
I waited for a while in case someone would ask the question for me. Here goes nothing. What's in the image?
there is a cigar in the wall
I knew about RCE using reduce, thats why i haven‘t solve it yet. Seems i need to forget about that😆
can see now thanks
CIFAR and MNIST, i'm sure its the same, the solution is right in front of our nose
Does the cifar counting challenge only involve "count"? The first two numbers in [125,245,0,10000] look like pixel values.
can't answer
Is it possible to do this with a model that was trained with private data ?
i think i'm finally on the right track for inversion
even if no flag yet 😦
i should cache all the answer i'm making to save time 😅
ouch Pixelated finally ! It was driving me crazy...
use lru_cache from functools 🙂
aha i'm always into overkilling stuff, i created a cache in a s3 bucket that I can load for all my ongoing sessions :p
i do, just lurking in discord from time to time
"Is it possible to do this with a model that was trained with private data " from papers googled, it seems yes
Been looking at this challenge.Surely Occam's razor must apply? Too much complexity would lead a contrived and arbitrary 'solution'. And this is not a hint since I'm nowhere near solving itAFAIK.
This is phishing 👆
When you join you're asked to verify your account, which sends you to a fake Discord login page
Or maybe it's a hint to Passphrase?
my mom always told me to never click on a link from someone i dont know
granny3: back on my 0.000693 limit. Let's see if my new method manage to break it
Anyone else having the feeling chall 27 not working as intended?^^
I thought I had a pretty good idea on what the chall wanted me to do, and half way there, it randomly spat out the flag on something that didn't make sense
"This is phishing"
soical engineering to steal our flag???
trying to hack the hackers....is this a test??? 😅
cc @tawny vault maybe? 😊
Looks like he's spamming it in every channel, it'll be removed eventually I assume
The competition just reached 1k participants 🙂 a lot of motivation around...
There is still plenty of time to reach top position even starting now.
I hope new participants will come in the next days
i image there is a few big competitors not thereyet
I just got flag for Granny 1 after 9 Days; Now I feel stupid.
Though there was great learning, learnt about different kinds of Adversarial attack.
got job, after granny one comes naturally granny 2
I think it's better if someone can add a discussion in the competition page mentioning that the discord is quite active for this comp. For me tbh i just was looking around here and found it to be active.
yeah it make the time pass faster while stuff are running on the side...
gosh i'm desperatly looking at my notebook to see a flag pop out of inversion
btw for the people who did inversion, we agree that the flag is also delivered as {"flag": ...}
just wasted 12 hours running queries against inversion with a tiny bug 
because I stopped printing the error message, i'll print only if "flag" is in the keys...
skill issue
no, it gives you another url, which now returns 184 outputs. jk, it is {"flag": ...} as usual
there is another question here , the static flag is flag{text} on only the text
¯_(ツ)_/¯
this also might change your queires
i hope it is just "text" otherwise i wasted the last 5 days of my life
Yeah, what? Say it isn't so
I'm assuming it definitely isn't
surely it's not a hint for someone who solved it so say so?
I dunno ... its a question in my head ! that is driving me crazy also
last year the format was flag{blabla} but for the other static flag (sloth) it is just the text so it should be the same here
i might have found a (new) piece of magic, but still don't manage to get the freaking flag
sloth static flag starts with flag{...} as cleared in the starter notebook
ah it is ?
oh you are right
I have to imagine that inversion isn't in that format, specifically because that it's mentioned for sloth in that notebook
gosh i hope its not the case for inversion 💀
there's a reason why it does
last year was a 6 letter word to guess, and still it would be input with the flag{...} no ? (i am not sure i wasnt there)
I think last year, the problem with same idea as inversion gave the flag instantly with no {}
it had the flag{...}
I mean in the images.
ah
@olive ledge can we get a clarification on this ? For inversion, do we need to query the static flag with flag{...} ?
its not really a clue, more a formating stuff i'd say
Obviously this is old. Follow the directions in the sample notebook.
ok thanks !
Granny2 solved. The hard work is starting now... (BTW: Funny easter egg)
What are the totally unsolved challenges? Granny3/Hush/CIFAR/Passphrase?
as far as we know yeah
So at least 3 (I don't believe much more) guys have solved Inversion
I believe it's at least 5
since 2 people not on 23 have claimed to have solved it
how can u guys guess all the 6 parameters of mean and std in granny
Yep slightly more than 3 ppl have it
I think it just resigns when it knows it's losing rather than playing until checkmate.
I'm still at pickle😭
samesies
most people seem to have gotten it. We just have to try harder 😄
I'm so stuck at Pickle that I think I have anger management issues now
This is a general good advice for most of the flags
pickle 🔥
I tried in pixelated things you wouldn't believe ... no flag ... but I feel I am just moving around it :/
I tried things you wouldn't believe in both of them, and got flag in neither...
hard work will pay off sooner or later (hopefully sooner for all of us )
I am focusing on hush now with a theory I am validating.
Nice. Good luck.
you too 🙂
tbh, this comment made me think that this problem is still more complicated than what I am trying lol
I wouldn't want to mislead you. it's just that among the ones I solved, this took me the most.
Many people already solved it.
I think my biggest issue with inversion rn is trying to determine which set of letters/numbers to use (uppercase/lowercase, yes/no numbers, etc). It's annoying given that any attempt requires at least four hours of brute-forcing because of the three missing bits of info and every different set returns different results...
i'm far from having solve it despite all the time spent on it, but today i felt I put the finger on something new
I hear that, I'll try and think about it differently. At this point I don't want to run another brute-force attempt until I'm reasonably confident about the five pieces of info we have
if it was pure brutforce, i think that by now, there would be more people having solved it 🙂
ez
a casual 67675234241018881 attempts needed
surely aws can handle that kind of demand
surely right?
right?
Absolutely
if you work at openai, you can maybe borrow some ressources
Somehow I feel it's either just lowercase or lowercase + numbers honestly, so only 2821109907456, seems doable
67 quadrillion requests?
nah surely the stopcodes are included :clueless:
hey everyone, i am new here. Till now, i solved 6 challenges. looking foward for solving more
welcome, good luck, and don't forget to eat and sleep
hello new here, i am kenyeung
there should be a warning somewhere about the danger of solving a CTF without sleeping
I know it's irrational but I believe I am getting close to passphrase
Currently I am stuck at cluster-1. tried so many things but won't stop
Also @olive ledge are we supposed to be able to batch requests to the inversion API? I always get 'Invalid input' with any size > 1. Would make my life much easier
is there a single unique phrase for Semantle2 or possible more than 1 asnwer :?
¯_(ツ)_/¯
... im going crazy
im starting to question if anyone has solved mnist
or if everyone is lying
its a conspiracy
do y'all have multiple notebooks open at any given time?
one kaggle, couple in google colab, lots of local ones
Yes solved, not conspiracy
the legends tell that the chosen one can solve MNIST and CIFAR with just an abacus
smart
can you pickle and resume entire notebook state? lol
The shape of my eyes currently doesn't make sense .... need to sleep 😅
The most difficult part about MNIST is that it's really easy to overthink it. It is by far the least satisfactory flag I got. First, think what is a reasonable explanation of what is laid in front of you in the notebook (ofc, w/o overthinking it). Then, do some trial and error, and you get it.
i think ive tried every reasonable approach possible
without giving too much away, ive tried every scenario that fits into the two values in the input_data list
I was in the same rabbit hole, go back to the first things you tried. Usually for this question, the first few things you tried can lead you to the solution, not the ones out of desperation.
inb4 next kaggle comp is predict which discord accounts are spam ones
any clue for pixelated? I keep staring at those numbers in days 🥲
{'negative': 0.3333, 'neutral': 0.3335, 'positive': 0.3332} still no flag ....
welcome to the club of 0.333 and no flag 😄
How dangerous I need to be☠️
name first 5 easy ones in there.. i'm sooo clulesss
I hate pixelated , I spent 2 hours trying to make the ocr work with the query then no flag :/ ....
check the discussions on the competition link on kaggle it should get you started
make sure to look around for resources here and on kaggle
good luck 🙂
All great points by Ali! I would add that if you're just getting started, don't stick to one challenge, but try a different one if you get stuck.
🥺
@brave briar any clues or suggestions or resources
Basically I cannot give you any hint :). But for this specific one, things I found on this channel generally misled me. I also tried chatGPT, that also make me lose lot of energy. So I would advise for this one to just search by yourself.
and also, I advise you to try to solve it locally, not on kaggle notebook.
what is the flag 1-6
I just solved pickle, I have no idea how, and I feel so empty 😄
or rather, i have no idea how it relates to the instructions
@amber glade Sorry to ping you on a Sunday, but just in case you hadn't seen this
@ornate marsh how are you affiliated with this and why did you choose to pick me up among others
I just saw you listed as a moderator in the side tab. I'm not affiliated with it, I was just reporting a scam. Sorry for bothering you
Thanks for letting me know @ornate marsh
@ruby hearth Please refrain from posting such messages here, I will have to take this message down for now
@fading flare Please do not spam this channel with messages like the one you posted inviting people to join other Discord channels (Crypto related), I have taken down the message
I spent too much time on this problem with no flag. Given what people wrote here i just feel that this problem is too random to deserve a lot of time tbh. So, i think I'll focus on the other problems.
The hush server is slow ? anyone tried ?
around 3 sec for test record.
hmmm
it took 37 secs to response for test :/
I checked from colab , kaggle .... best time I could get was around 22s response 😅 @olive ledge is the hush server still with us ?
yeah, I checked, same 20+ for me now, yesterday I had like 2.5s avg for 2k+ requests
yesterday was fast ... I guess its time for hush to scale 🙂
Not solved it and feel like I have tried everything "simple".
got many senteces with 0.33 0.33 0.33
never had a flag
i did that the first so i didnt though much about other problem didnt overthing
are you guys really have any hint what hush might be
i dont really have still discovering and mining
For CIFAR, I'm wondering if sample input data provided in [125, 245, 0, 10000] is real data from expected computation we've to submit or random numbers just for example
I think no
for the mnist problem, i checked the input data and it didnt match
I have an idea I want to test for hush , but the server response time is currently not so helpfull
I'm wondering the analogous thing for MNIST, but if it's just exemplar then it would be a stupid problem not a simple one
Not solved, but noted that "input_data is a clue".
Its indeed a clue
"For CIFAR, I'm wondering if sample input data provided in [125, 245, 0, 10000] is real data from expected computation we've to submit or random numbers just for example "
difinitely not a sample, but 10000 is also not a random number i think
"Come up with a difference sentence where everything is equivalent."
all score are 0.33, num of words =33, server runtime=33 msec
i just keep giving '=' and didnt get flag as well
did you reach 0.3333, 0.3333, 0.3333 ?
👍
example input and output in Passphrase are both somewhow important, as orgs mentioned
ye
I'm still stuck at cluster1 any help please?
No clues, but there is more than one way to solve either using the provided model or otherwise.
Asking for hint buddy
Competitors can't giving be giving spoiler hints.
actually thats waht i did then , then after solving other i got to look and cluster 1 and understand what is was doing
I'm stuck on some other too 😂 like cifar
All of us are stuck somewhere.
😂
@bold ibex
I'm asking for a hint not for a spoiler
Solved?
if yes i can take 2 more days to solve it and i will also reach the 23 zone
Were you able to get Granny 3 in the local model ?
not solved yet, but getting frustratingly close
Did all the what the flags but 3, that childish response is driving me insane, cant seem tobypass it 😭
cannot answer about what i'm trying 🤐 but still stuck at 0.00069273, so i'm going to explore new horizons now
For passphrase to me différence may be related to some mathematical logic
you: going to explore new horizons in granny3
new horizons: 0.00069274
any hints on WITF, have tried simple to difficult prompt, it says nothing useful
Ya im lost if WITF 3+ lol. Just jumping around different problems and failing at each one
I had success on 4,5 with the same logic for 1 and 2
3 however.. not budging at all
For witf there is a very helpful yt video that even links to a website with similar competitive tasks that teach you a lot. After watching them, i've cleared all the LLM tasks within an hour except IP and pickle. And the topic is very researched and researchable
Furthermore, clearing witf 1-3 with a single prompt
It is indeed interesting whether you can clear all from 1 to 6 with one prompt
I'd tell you what other techniques I used, but I won't because competition rules
Except this gif
the one i found the hardest has been the pirate that talk about flags
that is a very helpful tip, will look into it, thanks!
Trying to get to 10 flag club today
https://tensortrust.ai/ - there is a game where you can practice attacks & defences on LLMs 🙂
Rise to the top of the Tensor Trust leaderboard by fooling AI language models, and help researchers make more secure AI along the way.
MyDay :
while True:
for task in ['hush', 'passphrase', 'granny3', 'inversion', 'cifar']:
continue
now I still 2 scores T.T
i wish iwas like this for thi challenge
Can you share it??
i doubt that
this story reminds me of "Good Will Hunting" film
Sleep is all you need!
even harder than IP1 and 2?
btw, is there any information what model is MobileNetV2 is ? it is Keras or torch?
I mean is it announced by host? or we should figure out by oursevles
for ip1 and 2, are the inputs dns records or prompts like wtf1-6?
no idea, did not solve them yet, but it says: "We have a service where you can query it in natural language and find out where to send your traffic",
Need to figure out yourself.
@minor falcon Yeah solved all witf with a single prompt but pirate is taking too much time how should I approach it. What's the difference between pirate and other witfs. After some days I started giving prompts in pirate accents even that didn't work...😪
Edit I got the pirate flag
did you got IP1 and 2 ?
I got IP1 after they modified the problem statement
solved granny2?
Yeah
are the inputs dns records or prompts injections like wtf1-6?
4 days on granny, can't still figure out anything...feeling really dumb
figuring out expected input format is part of the challenge, if it's not clear from the problem statement
i was like 5 or 6 days on granny, dont give up!
really difficult not to, I'll try
I think you have to experiment.
thanks, i just solved ip1
pixelated is teaching me a lot (really a lot) in the security world ! -- still no flag .... but much info gained
interesting I found pixalated quite easy even though I knew nothing about that stuff.... but mnist 🤯
hhhhh I have the opposite answer completely
but pixalated gives you some feedback, mnist is either you get it or not there's no progression
ok just solved mnist
LOL
kind of hate myself as I had this idea but didn't bother to try and now I fell outside of the bronze band .......
you may want to cancel this comment
how hard did you guys find sloth?
I have a very logical assumption of what I should do in the task
given the values
BUT
the output shape doesnt fit
has it been solved?
yep
I pretty much just found the staticflag, just need to read it 😅
Working all day on sloth because I thought it would be easy but no luck. Went back to MNIST to try 1 thing and finally got the flag 😅
ive been on mnist and cifar for 3 days+
is hush server down?
Checking now
granny is also slow. It used to be faster. It was the same last week end and during work days, servers are more available 😀.
The hush server is down or very slow its been so most of the time today ( like 12 hours so far 😅)
I can’t really test
Working on it now
Great thanks 😊
huh took me an hour, back to bronze
You solved inversion ?
nah haven't tried that
Ah static you mean the sloth
Services should be back up and re-scaling.
My bad 🙂 , once I see the word static I go inversioooon ! ....
almost 6h in what is the flag 3, this one is not budging 😔
even pirate only took 3 prompts, by witf3 still wont budge
you'll get it 🙂
looks like your prediction was wrong :p
not a bad thing to be wrong about
depends for who 😭
Still stuck on pixelated … I tried everything I found or learned about and I know why it’s not working …. I am overthinking 🙄🙄
for the mnist problem, is the second part of the input data really not relevant??
try it either way
How dangerous I need to be to get the flag😔
you need to be perfectly dangerous
many have claimed they could solved MNIST without making sense of the 23435 ... which looks magical to me !!
there are 23435 possible solutions
Ip1 and 2 done, still no witf3, going insane
is 403 errors (I guess because of too many calls) shared between different questions? e.g. will I get 403 more often if I do a lot of calls to granny and pixelated at the same time compared to if I only do granny or pixelated?
when i trigger the 403 i wait 3min and regen the request, that works fine so far
Yeah have kinda same workaround, just seemed to trigger more often when I started my other stuff on the other questions simultaneously, but might just have been a coincidence
Is semantle2 possible without bruteforcing?
possible or, has it been done like that so far?
i believe i've won
I didn't brute-force it, but it would've definitely been easier if I did
like many questions, it was a combination of both; either I start with bruteforcing and use intuition for the rest, or I start with a hunch and exhaust the options
Got them (minus #3) with a single prompt. Will be fun to look over and share these after the competition ends
so your solution does need the magic number 23435?
count could mean say number one after another or calculate total number of sth. tried a number of ideas, none worked.
I don't get how people have solved inversion easily, there must be a "best" way to do it to get everything to line up and it's definitely not traditional model inversion
LLMs are like that.
I think it has to do with the datasets you happen to use.
I am also banging my head against MNIST... I've found a few different ways to arrive at 23435 each of which could arguably be the solution, none of them seem to be 🤔
Hmm, I'll rethink my strategy. Also was wondering re my question earlier--is batching supposed to be allowed for the endpoint? Every time I restart my notebook I need to send another 1000-10000 requests to build the train/test set for my model. Would be much easier to just do it once/a few times. (Guess I could just save the output actually)
finally got MNIST after 8 days of ovethinking 🙂
You technically don’t need a notebook…
We’re discussing making some changes. They’ll be announced well ahead of time - but it’s a competition so we’re generally on the “no changes” train.
I’d rather people are mad at us for a misguided implementation, rather than causing someone to lose a spot.
Makes sense, honestly I don't think that changes really need to be made, maybe a few things could be made more clear, especially with Passphrase (as I think the main difficulty is actually figuring out what to do), but by-and-large the competition has been great thus far (as long as granny3 is solveable 🙏 )
No one like to lose a spot 😅
nice, solved these three yesterday, same prompt worked for me for ip1 and 1p2 😄
i am still stuck at pickle ... i wonder:
- if my payload is well concealed, it will not be detectd and will be unpickled.
- if my payload is detected, it will not be unpickled.
Or there is no unpickling at the server at all???
Exactly the same issue here, 3 won't budge at all
too much info, won't say
you can assume both and go with your assumption
@random minnow I must say I enjoy every thread and discussion you are involved in, A researcher by heart ! 🫡
not sure if it is okay to ask or not, but can someone clarify if inversion static flag makes sense? stuck on it trying to make sense of output since last 3 days
I didn't solve it by far, but tbh i suspect it's a LLM problem but can't find evidence for that.
I'm pretty confident that it should make sense but not 100% sure. Also not sure if it's a word or phrase
I also think this information should not be given to us. The fact that it was possible to solve Inversion without having that piece is enough for me.
aka it would be unfair to the people who solved it this far
btw, knowing that something is solved is also information
f.e. if some search space is obviously too large for someone to just get lucky while brute-forcing, you know there has to be a "trick" they applied, some extra knowledge/method etc
and knowing what is "obviously too large" is part of the skills one should have to do well here
....say I while trying to brute-force Granny 1 🤣
I wish I had a solution ... but I didn't manage to understand the clue yet
who have scored above 10/12, how many days you guys devoted on an average?
got around 8-9 in 1 day, 12-13 on the second one
mostly on LLM things because they are rather formulaic and similar in the approach
Yeah agreed, LLMs, Cluster1-3, Semantle1 go fast and that's 15 right there. However for reference I've spent around a week on Inversion. It and Grannys have taken up most of my time thus far.
moohax — 15.10.2023 20:18
There is no actual code execution.
oh
i'm giving up on inversion for now i ran out of ideas, and even if i thought i figured out the magic, apparently not 😢
That's where the social engineering part comes in
How much, it’s time 😭
Alright
I will hint you
The hint is: reread the flag prompt
please don't make me go back to this nightmare aha
Like honestly, looking back, it might be helpful to try interpret the prompt in some way
Except what I can reassure you of is that Q in Ouija is accidental and does not hint to anything (atleast within this flag)
making progress on your remaining tasks you ?
Nope, was resting on the weekends, although I have had a few thoughts on some of them and may try them later
is it official that the Q is accidental? 😄
nope. Like i said, social engineering is a part of the competition
...and now you'll doubt, whether you should believe me or not
you're playing 7D-chess, I'm struggling with hungry hungry hippos 😄
It is official, actually.
french version, which is also with a "O" according to that same discussion 😅
Nope, MNIST and inversion are driving me crazy 😩
mnist is manageable, especially with all that has been said in this discord
I will re-read everything, sure there's something I'm missing
some people really know how to make money out of nowhere 😭
the only thing that has been said is "dont overthink it" 😂
when joe@simpson.com will reply to my mail?
maybe counting mnist needs label
did you send your kindest wishes or warmest regards?
is it weird that I was able to solve Granny 1 but not Granny 2?
Literal skill issue, lol
🤷
the inversion competition🤔
i'll be so pissed when i'll see the solution revealed and discover i was off one character 🥺
ah, a new guy has got to 23, congrats to him
new ticket for itm going closer and closer to 24 flags
i spent one more day to solve granny2 after granny1, so it's ok
making progress Mikhail on the golden square (the 4 apparently unsolved challenges*)
at least you solved one
not much yet, researching hush, tried some ideas on passphrase but execution is hard and given vague prompt it's probably not worth it.
my new ideas for cifar immediately fall to "way too overthought" category.
granny3 is ☠️
yeah CIFAR everytime i find an idea that would fit in the "not too overthought" i end up with the same rewarding message "try again"
i'm wondering which one will be the first of the 4 to fall the first
I took me literally 2 queries for MNIST, but CIFAR is something completely different. Maybe someone from club 23 solved CIFAR, we cant be 100% sure that's these are the same 23 as mine.
yeah maybe
cifar is really vague, but the fact that it is so vague is maybe also a strong clue
I haven't looked at hush at all because audio data is intimidating but I at least have approaches that I want to look at for the other three. I think at least CIFAR and passphrase will be solved by the end of comp
yeah me i keep the hush for after inversion. If im not able to solve one, i don't see myself going to the other
passphrase, i'm not sure someone will be able to solve it because
- its wether a trick question and we are all missing the subtility
- there is a bit of brutforcing (or at least smart searching) involved, and the indication is too vague so its hard to see if we are heading toward the good interpretation.
i'll bet more on CIFAR/Granny3 to be solved, at least thats where i'm currently putting my efforts
regarding passphrase is there a guaranty that if we got:
0.3333, 0.3333, 0.3333
it will give us flag?
nope, thats what i'm saying, nobody knows what is exactly the objective, and there is many way of interepreting the message given (without making a list here as it would give too much info i guess)
I think we might get a hint if it stays unsolved for much longer just because the difficulty is in figuring out what we need to do. I'm at the point where I can influence pretty much any three pos/neg/neu values to 3 sigfigs but it's all a moot point if there's a single solution or something
i definitely prefer the challenges where they're harder but you at least know what your goal is (more like semantle 2)
or sloth even i think that one was reasonable
I'm leaning towards the passphrase solution to be semantle3 though, noticing some weird trends
might be also granny - related. Maybe a certain score is to be reached under certain circumstences, and the flag would not trigger overwise
In chat you can find examples with 0.3333 all and no flag. And given that "both example input and output are important", that is probably not the way.
this is one hint about passphrase
hmmm, I did see scores like up to third decimal point, but not exactly 0.3333, thats why I mentioned that, maybe the goal is to be super precise!
there were 100% perfect score posted, indeed
oh btw, did you got passphrase ?
like there was also almost perfect score posted for granny
ah didnt see that one, interesting.
and that the second. but maybe it's not about Passphrase 😄
afaik passphrase is unsolved
going toward the granny observation. We are maybe all trying to submit a picture of an apple while we should be slightly more subtle
For Iversion, the prompt says "It's basically an AI Quija board.". Is is supposed to be "Ouija" board? Cant' find anythign about Quija board
yep thats a typo, it has been clarified by the organizer
Thank you!
@empty bane how is your progress? still struggling with inversion or trying other stuff?
Inversion mostly. But I've sort of run out of ideas, lost a lot of steam
Wow, first 24 solved on LB!
Trying to tackle Inversion, Is "static flag" as in Inversion also a concept used in last year's competition? I am guessing the static flag to be a 8 char string, but I am begining to question myself.
Who is Bingus Traveller The love doctor
thats ashame this person is not here on discord, would be curious to know which flag felt
but at the same time the fact that other people dont know is certainly a hedge for him
i was thinking that, trying to spur us on
Maybe it's the main reason he got 24 first 🙂 unbiased ideas and more time w/o shitposting xd
but no, they participated in a norweigan university internal competition 2 years ago 👀
without shitposting, i'd have abandonned for a long time personnally aha
I think it's definitely CIFAR that he solved, I really doubt any of the other three
@olive ledge can we get the statistics of how many solved each problem
even if I wish we could have them, i think its a bad idea, it would penalize the people that are alone having solved a particular problem
like if i'm the only one solving granny3, i will certainly not tell it, fearing I lead everybody to focus on that one knowing its more managable than another
why i end up with teapot?
according to this rank, sloth is ranked as easy as IP1,2 and Cluster1,3. Does this makes sense?... Saw someone spent days on sloth...
it starts with "based on my experience", so it depends on person,
myself I found Semantle2 is much easier thanCluster1, which I'm stuck in now
Probably a very personal ranking, i got IPs within minutes, and have 5h+ of cluster1 without any flag
We all have kind of nemesis which seems easy to others 😀 . Due to skill ? luck ? overthinking ?? Don't know. It can be MNIST, sloth, granny, pixelated, WTF3 or pickle depending on each one.
Are you professional semantle player?
Right, I see....... for me, almost all problems with directional feedback are quite easy
Don't give up, it took me 2 days to get 19 flags, then I went stuck in MNIST for 8 days in a row due overthinking 🤯 . Now in a couple of hours got Granny1&2.
The ones without clear directional feedback are just so frustrating...
hmm no can't say it
At 13 flags!
This is when you are too good at Kaggling 😄
passphrase is just hard to understand, doing analogies that we are supposed to do make things even more difficult haha
god i hate pixelated
the task itself is awesome but the OCR makes me wanna pull hair off of my head
I propose each person to get a joker to use for one of the task 😀
how sometimes letters just disappear
actually i did not find the OCR to be so bad
i struggeled a lot with a particular character, but it turned out i did not need it...
its like 95% about the font and the way you alter your picture and submit it
don't overthink, great ressources for this one in the conversation. I got stuck too but it helped
old london is all you need
but even after excluding the font and minimizing other factors i still change 1 pixel to 254 254 254 and the whole 100 character query goes wild
I don't even know which font i used
for the reference, old london is one of the first font i tryied for inversion 😭
old london is legendary
funnily enough i did find them which SEVERELY helped with the consistency
or something like Gothic
you should try different approaches to generate texts
(worth for all the people stuck with OCR)
but still i've spend 2 hours failing to make ocr recognize one last mislabeled character correctly
like i'm pretty sure that like a few more alterations and i'll manage to make it but still that is really maddening
out of like (idk if that number is too much of a hint) 50+
yeaaaaa i was certain that Paint, despite helping me with some other flags, may not be the best tool for the task
guys you are saying too much ! I know it s hard 😀
if you are talking about passphrase we are only relying what moderators told us 🙂
no sorry my first message was a mistaken reply, I was talking about OCR talk
oh np
let's switch to granny hints 😄
agree
let's share inversion hints
staring into nothingness🥲
https://www.kaggle.com/code/moohax/ctf-starter-aiv
shared mine, and yours?
alright actual inversion hint
the flag message does not contain unprintable symbols like \n \r \b \a \0 etc.
inversion hint: the flag is the name of the last person who solved inversion
i give you a nice one liner for inversion:
while True:query("".join(list(np.random.choice(list("abcd...."), 8))))
Am i the only one who is receiving primate messages from anonymous accounts to share hints ?
I haven't been asked to share hints yet... feels bad
few people have mentioned that, i havent received them myself
but also if they're in this server then they're not anonymous
My count is 4, 1 is even active sometimes in this chat.
i got one guy that tried two times, i threat him to talk to orgs, he stopped.
you can create a new kaggle account/new discord account, and you are anonymous again
No one asked me 😢
you are not shitposting enough
well now you can be sure somebody's gonna ask you haha
that's true, most cheaters don't think it through tho
if they could bypass anticheat they could solve some flags
am I the only one ending up with 100 notebooks starting like notebookxxxxx or ctfxxxx and not able to find something when i need it ?
I have 30 notebooks for granny
only in my terminal
just 3 for me
I have only 3
my all Untitled notebooks on colab are related to this 🙂 ... now its Untitled47
I am to lazy to name it :/
But once the interactive mode was alive for 27 hours 😪
oh one thing .. if the UntitledX with a flag , its named UntitledX_yes 😄
i only have one notebook and i delete ctf parts that i've already solved
I just quick save them and delete that particular problem from the notebook. Gives a big relief ☺️
yeah exactly same
afaik for top5 you need a notebook with solutions and a write-up so i'm not deleting none of my cells
Lol... just got the joke, I am also sadly getting a teapot
if someone asks you in private for some hint, give him a very bad clue that will give him no chance to find it out 😁 .
teapot and coffee?
Now I am confused whether the teapot is a standard implementation/joke or actual hint..
yeah my inversion folder is starting to get depressing
what I am inserting to pixelated makes perfect sense! yet no flag! I started to like the sloth because of pixelated 😦
omg ! I did it pixelated done !
stupid silly OCR
22 flags finally 😄
Are there any predictions on how many will be needed for bronze in the end?
I think we finished it at the same time, I was about to come here and vent a bit about it but my solution worked 😄 Congrats bro 🔥
Trying to solve Inversion/Granny 3, valley of despair it is.
Does the order of the words matter in semantle2? 🤔
Edit: does not seem it does
I am already working and focusing on inversion (8 or 9 days so far) and hush ( 7 days I think) ... so I know the feeling , but pixelated was the one that drove me crazy because I know I had it !
For inversion, the AI want us to guess 457 as it never learns about it or it's another secret to figure out?
Are there consensus yet whether pickle is an LLM?
yes
i think the answer to this question would give a free flag to 99% of the people stuck 😅
It definitely matters for a random sentence, unclear if it matters for the solution
And the consensus is that it is ineed an LLM?
¯_(ツ)_/¯
currently the consensus is focussed on other tasks 😭
dude im gonna spend the rest of the competition on mnist i swear to god
ok i'll give you a hint to solve mnist you need to
ah wait someone is ringing at my door...
Is it moohax ?
hmm, i didnt have the same experience
i didnt feel like the ocr was that bad I got a hang of it
Interesting, dont have the answer yet, but ran like 4/5 tests on some words and it provided the same score 🤔
thats interesting
got semantle2, I guess I need to turn my attention to granny now
your thoughts on semantle 2, like it? i stuck at 0.98
Im at 0.89 on sem2 🥲 so close, yet so far
I guess just keep going I didn't do anything smart at all
you solved granny 1 and 2 then?
yep, they are actually not that hard
hopefully 😅
Pickle is so frustrating.... Many suggested to read the prompt.. "This is not the code execution you are looking for...waves keyboard" yet there are just so many possible interpretations...
Is the pickle flag static?... Feeling like I am playing hide and seek with the notebook.... Or is that a waste of time...
well well
I've a few (different) perfect matches but it goes nowhere to me
yeah i'm at the same point, i also thought i figure out more, but it still leading nowhere
finally got the sloth but can not get over the pickle
F1 score of the underlying model might be quite bad 🙂
the bad part is that some characters can be easily swapped between upper/lower case, so its multiplying the search space if once want to include everything
there is a 24 on the board already 🙂
Ya I know , I meant that 23 now is not enough 😉
ahh dam, best of luck those going for gold haha
Stuck with MNIST. Really confused if it’s related to the actual dataset. Seems like the tensorflow one and others aren’t same. Which one to use?
¯_(ツ)_/¯
i'm getting tired of 'I know you are, but what am I?'
humans generally get tired before machines 🙃 . Take a break and fight him again ! Definitely the hardest of the 6 to me !
I know you are, but what am I?
tensorflow is same as pytorch
atleast from what ive seen
but ye i am also stuck on mnist
tfw a prompt for witf 1 accidentally worked for witf 2 and 3 and you don't even know their gimmicks
no one has solved granny3, hush, and passphrase afawk still?
oh, so the one that fell today was cifar?
someone has 24 now so one of the 4 is solved
then theres me who solved 4 witfs with the same prompt 😂
yeah exactly, lots of silence here regarding that haha
people hiding the one that fell 👀
nobody knows here i'm afraid, we are all busy figuring out our own troubles
fair fair, jsut really curious
I have got all 6 😂
starts overthinking
arf, i thought i had the illumination. Found a match that would fit almost perfectly with all my findings... but nope.
many such cases
You can search over this channel to find a hint about datasets to use in mnist
I'm giving up on granny...can't match the models 😔
having trouble understanding what kind of output format we are suposed to feed to granny
Look at the starter notebook
ok just got mnist and i can say i totally didnt get the input data clue
brute force?
You mean the two numbers in the input_data is not needed at all?
¯_(ツ)_/¯
im saying that i misunderstood the clue
okay.... yeh, we all do until the aha moment.
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.
in pickle, do we seriously need to be dangerous or it's just another LLM?
sharing your pain...
dang.. just got sloth. One of my favourite so far, one that you can do with pure reasoning and not guessing/word play
sloth seemed too easy compared to the last time
Yeh, multiple clues manifest itself and point to the same direction, no guessing involved, just so straight forward compared to others..
I am crying because people are saying same thing about MNIST as well and I am nowhere near
same here. The input shape looks like dead giveaway and stragith forward, except no flag
switching to granny 1, seems doable. but I heard people having perfect score but no flag? I thought I was just supposed to upload an apple, apparently it didn't work
So you are saying that I won't get flag even if I reach 1.0 for granny ?
I missed the word "almost", #ai-village-capture-the-flag-defcon31 message
Or maybe I was misreading the context
But seriously I don't understand the goal, my first thought is to get a socre higher than any other class would be fine (come on! it is level 1!), I was terribly wrong..
Yeah I did the same I also tried to keep the timber in the second place and granny in the first that didn't work too..
Lol, I did that too, try to keep the wolf around, and still no flag
The other hypothesis (which I didn't try) is to put the wolf at the bottom, but that would make no sense...
It's possible to do that I'll give it a try once
Also why did they name misclassifications question as cluster 1. I thought it will be easier than 2 n 3 as I solved both in 20 mins. I don't know what to do ? Any resources to learn about it?
cluster 1? Just look at the data, and do real data analysis. Straight forward, no word play. If stuck, just try and let the feedback guide you.
But you are right about the ranking, cluster 2 is a joke to me...
I've heard in the chat somewhere that you can get the flag for cluster1 without using model,
Is it true?
I think that's true for multiple flags.....
Well, maybe. Try it
Well, MobileNet-v2 mentioned in the prompt scores ~85% timber wolf for the pic provided, and not ~28% returned by the query. So the model/prep is obviously different and people are trying to match 🙂
Are there any good tools to convert text to img?.... Using MS paint is not working well...
Python?
Good suggestion, when I searched text to imgage, I got all Generative AI results.. Need to work on my google skill
stucked in cluster3 more than a week. response messages for the same input is constantly varying alphanumeric string of length 21. is there any meaning? or just by chance?
Sometimes you have to just stare at what you are getting
ok seems a different angle, think about it doing semantle2
OCR really sucks.... funny that there is a font called OCR and it is terrible at being recognized
Chatgpt told me that the response I got is typically an indication of successful insertion. Yet not flag, sigh....
waste 6 hrs in IP
I am sure there will be memes about IP1 and IP2. I got the two flags and I still have no idea how...
I have no clue now on IP.
You just need to try, and at least you know your input are being processed properly when doing the IP. The OCR for pixelated is just wasting time
yeah.
one whole list of papers: https://github.com/stratosphereips/awesome-ml-privacy-attacks
how do I interpret the score in cluster 1? do I maximise it to get the flag or minimize it?
Make hypothesis, and then test. This one should not be hard to test.
in granny, the local model can only match the server up to 7digit
is it a match?
or i need the exact number
another 24
cool it will be time soon to give up and be free :p
pickle is really testing my anger control at this point 🥲
same. lastly found some inconsistent behavior which I deem too random to be part of the chall but which makes me question WHAT EVEN EXACTLY IS THE CHALL.
Thought it had a very straightforward goal and I just had to try hard enough, not so sure anymore
same. cluster 3, pickle and pixelated really confused me because I think I found many clues in all three of them, but I tried for days and couldn't get the flag. sad...
even for granny 3?
I dont think i have any clue at all for pickle, reading on base 64 and serialization, some ideas come to mind but nothing seems to work
i wake up and sudden;y solved pickel????
just reoragnize my code and reimplement old initial ideas
wasted one week
congrats! 🙂
Now it's inversion time
the two most solvable ones in my opinon are inversion aand passphrase
both are model inversion which can be done
for passpharse, i am worling towrads this hint "The sentiment about this sauce became so powerful that it was encapsulated into an AI model."
isnt passphrase one of those that ahs not been solved yet?
Which are the problems still unsolved?
there is a chance that everything is solved
4 days later, pixelated's ocr recognizes my text correctly. the issue is that query is way off :/
3 people in last 30 minutes 🤔
yep also there is sharing behind the scene
🤔
I am not making any speculation.
I only care about the tip🤣
Hm
But i hope there is no tip.
ahaha
just good old team work
AHEM
hackers sometimes collaborate🤣
1 more guy to 23 from Italy
teaming with people you knew and sending flags almost at the same time 🤣
I guess they don't know collaboration is not allowed
@olive ledge ^
if they send the same tasks at the same time
when it is obvious teaming
99% sure they've waited for top5 to fill up to not do write-ups later
23 is not enough for top5 imo
gold medal farming
i think even my first 23 won't be enough for gold..
Im geniully naive but i think people would not risk to cheat, especially at that high level, and without any subtility. I'll vouch for coincidence here
The speed 23 is appearing is crazy 🙄🙄🙄🙄
Btw, this is my first active kaggle comp, do people below bronze get any points or something? or is it just bronze+?
You get a chocolate medal
you'll get a granny smith apple
Idk I'm a little suspicious, given how few people have solved inversion, that they seem to know each other, and they all got it within a very short timeframe
I can only think of inversion static key is shared somehow maybe !
mnhhh chocolate medal
same but they were in the same team before
Time to search in Italian socail media ?
Italian kaggle twitter scrapping
maybe this is some kind of a clue🧐
When I see the italian team submitting all at the time with so little discretion, I really wonder why to bother to respect rules...
Maybe guy at the first place is actually a team of students somewhere
10k$+ is not a bad sum
I don't think it's worth worrying about right now--we should give them the benefit of the doubt and trust the moderators to investigate as necessary. Lets get back to solving the challenges
I could have let the benefit of doubt for 2 challengers indeed, but 4 ... Come on !
I mean if they're cheating, it's obvious enough that I'm sure they'll be removed. No point speculating tho
Lets go back to shitpost* solving puzzle takes too long and too much energy
where are y'all stuck right now? I'm jumping between MNIST and cluster1 :/
But they did solve inversion, that's strong
Still inversion, life is pain
wait, how do you know that?
trying inversion, but seems I have difficulties to interpret spirits
I hope there is no cheating , a gold or silver medal is important for me … to push my way to master 😅
do you think they solved hush granny3 cifar or passphrase?
Inversion tip: you can only solve it at night during a thunderstorm
I don't know 😄 could be any, but yes, inversion is the most likely probably, I just thought that maybe they had changed their team names or something 😛
yeah all four named "solved cifar" 😉
I got mnist but I'm running blank on sloth..
LMAO, after lecturing here yesterday about the importance of error checking and logging in your scripts, it turns out I had not done it properly for my Granny 1 notebook and I lost my flag 😅
luckily got there again after fixing and rerunning it....
There is another Jacopo in that team, it seems 😛
so not even having the same name was enough...
It would be too obvious if two people with same name reached 23 together.
try submit your name to passphrase
I guess the minimum requirement was to have granny 1 and 2 solved but I'm too dumb
I did in pixelated, I got "you are not invited"
Bingus Traveller is another italian for sure.."love doctor", come on
moderators may have every log to check if they suspect cheating
they are all competing to another competition : NeurIPS 2023 and have teamed up on a previous competition : RSNA Screening Mammography Breast Cancer Detection.
You are right, I just went through profile of all 5 of them and all are competing in Defcon and NeurIPS individually.
But more generally, I wonder if this competition would not be even more interesting if team were allowed. Competing 1 month without discussing concretely of the problem is so hard.
Team is not a good idea here because you can split the tasks
what's even more funny...they are colleagues
at Leonardo
if teaming was allowed problems would be 3 times more difficult
dayum got pickle
agreed, CTFs are just so much more fun with a team.
More social, less being stuck.
Maybe make the competition only half as long but allow teams next time
when you did something wrong in front of CTF hackers and in 5 minutes they know who you are, where you live and even your cat name 🙂
Noted - thanks for the heads up. Will get this to Kaggle and we’ll investigate.
same, cluster 1 seems easy and still not sucess
indeed. Can't wait to see what I was missing 😄 Same for MNIST
This is my first ever CTF and I'm stuck on Cluster Level 1 
it is not evident but difficulty is not in order. Have you tried cluster 2 ? It may inspire you for cluster 1 😀
We are doing the challenge.... i can share my solution with the organizer if there are any problem @olive ledge
the other can do the same
i don't have nothing to hide
I think the organizer have the submission time and statistics and they can also check that
Sorry but ask for our solution if there are issues...
i can give mine immediatly if you need it
After that ill stop immediately the challenge if tehre are problems
if it is not a secret what flag did u send last

than i ll say whats the problem
is this a hint really? or you just try to motivate him 😄
I think the issue is that they assume that you guys are working together (which is prohibited), nothing with the solution itself
well an accusation is more than an assumption
im doing this competition for fun
and this is not fun
🧂
But are you guys working together? You mentioned “our solution” and seem to be familiar with the others
everybody has a own solution
btw my last flag is pickle, 22nd
so i can say our solutions. I think it is correct in english
@wind ether
what im missing?
Sorry, misunderstanding on my end. I was just trying to clarify the situation, not make any assumptions.
Anyway i wrote privately to the organizer because for me it is bad and i think ill quit this competition
You said everybody has a own solution. But that's not even the key question. The key question is "Did you work together in terms of sharing information privately"
The issue is that sharing methods, hints ... even if it is not complete solution is forbidden ! Everybody puts lot of effort to respect that on this discord. And in some way, if we all do that for fun, let's share all methods to all learn together.
I have the feeling that hush was faster on previous days
was very slow on weekends
Well fine i selected my first submission... now im 4 and ill quit with this competition. So you are fine and happy...
See you everybody
🤨
Peolple keep saying don't overthink MNIST, but if I have no clue about the two numbers in input_data (well, tried a few obvious ideas that didn't work), am I overthinking or underthinking?
I have stopped thinking about the MNIST and Pickle.
Pickle is fine, I have not solve it. But it pushed me to learn something new. MNIST is no fun at all.
Eager to know the solution to Pickle when the competition ends, seems fun, but out of reach currently
Luigi here 🙋 I am sorry for your concern. It is true that we know each other, and we work for the same organization, but I don't think this was forbidden by the rules. We enthusiastically talked about the challenge because it is fun and... well, challenging. This is exactly how it is done here. We did not share more information than what has already been discussed on this channel. @olive ledge
Wow, I just realized what needs to be done for Passphrase during a shower and very confident about it! Next is just implementation. Encourage everybody to take showers when stuck...
aight, running to the shower
This is indeed not a teaming event as stated in the rules.
We (and Kaggle) understand this is a different type of competition for many of you, some confusion is to be expected. However, the integrity of competition is kind of everything, especially if we want to be invited back next year. All participants are responsible for reading and understanding the rules.
Please know that not only are there post-processing checks (the details of which will remain private), winners MUST also publish their solutions to be eligible for prizes.
That being said, we’ll bring this up with Kaggle and follow their guidance.
Please continue to CTF! ❤️
Well, keep me posted, I'd be glad to stop wasting my life on passphrase if that's not necessary 🥲
No, please continue!
I suggest that all gold solutions should publish a write-up.
I'm impressed by how trick is this competition, even a single word or a gesture exchanged in a circle of friends can be the key to solving a problem.
@violet trellis it doesn't solve the problem...
"I spent couple of days thinking about MNIST then I realized that blablanla" how can you spot a cheater?
Yeah in this competition it is really hard to do that as almost everybody have similar solution.
in this particular setup, the only thing you can do is to check the timestamps of submissions
Wait to see my writting with all i tried for inversion ahah
No solution for this of course. This is just better than doing it for the prizes only.
One guy actually deleted their account.
Continuing to talk about this is probably not productive. We will talk in the backend.
Also, there are still community guidelines to follow. Please ensure you are following those!
Maybe hush requires an italian accent
Finally got MNIST. Now i wanna kick myself for not trying this first
'Don't overthink it' sounds so true now
this gives me hope 😄
Was stuck at 18 for 7 days. Got 5 in two days. I followed the best hint ever given : get some sleep.
me: trying to sleep
my brain: lets try this idea for CIFAR

Please stop offending Italy and Italian, we are NOT in the same team, but against each other and having fun, we will share the codes, ideas and solutions without hide anything, and we are not in this competition to gain money, but to learn and having fun, like all the other kaggle's competitions. And btw, who can assure that N people are not collaborating together and creating a single fake account appearing as one ? If we were incorrect we would have done so...
Now hurry up! have fun and find the flags 🇮🇹 !
Please stop guys, stay in the comp mind 🙂 It should be mostly fun for everyone
yeah and please share what u discussed between each other
especially about inversion
getting some weird outputs that are neither the flag nor scores in semantle2 🤨
ok for me

I dont think they mean anything tho haha, just unexpected
running out of words to try, and the dictionary starting to come up with all sort of names.... lol
Man ofc we all belive you it is coincidance that 4 friends put solution (for probably same hard task) in 1 hour period and all get 23 points 🙂
Don't make me come in here....I'll make Granny 3 even harder.
@olive ledge is the test_noise.wav somehow important for Hush, or it's just an example? (answer like "can tell" will be ok, but maybe that info won't be much of a hint)
It's an example.
i feel the pain in that "mostly"
sometimes Failed to process input pops up. Failed my search just now
ah, it was exactly that haha
with very different inputs, so got confused
Holy *, bathroom thoughts verified. Passphrase flag after 3 hours
another 24? 😮
do u need to be a native english speaker for that?😅
Great ! Congrats ! I have also a plan for Passphrase once I am done with the current work on hush
I am no way english native.
then it's cool because understanding the text of the task is a separate task indeed
embarassingly... I only have 18 flags so far...
I know it's cliche, but ...read the prompt
u will catch up eventually
Wow, getting Passphrase is impressive, congratz on this @gusty warren ! Think I gonna print the prompt and take shower with it 😄
i tried to register the prompt and fall alseep while listening it in loop, did not work.
Pretty much just look at what they did. First one decided to quit because he was 'wrongly' accused. Second one decided to delete his Kaggle account completely. FInally, this guy calls people out for attacking Italy LOL.
Probably best to just leave it to the organisers for now and stop speculating on it
I don't think we need to discuss that as for now, because it's always a possibility of wrong accusations and without strong proofs it's not correct to write stuff like that, because it could be very harmful experience for other people.
let's celebrate passphrase is being solved
has been known being solved* maybe other people already solved it without bragging :p
sure and let other people with high scores know that they can inflate the leaderboard with alt accounts and friends' accounts
yeah, disclosure of hard tasks being solved can be bad for you leaderboard place but surely gives hope and motivation to others 🙂
True. I probably woldn't talk about it if I were on top of the LB. Getting the passphrase won't get me gold since I started late, but may stir up the LB a bit by letting people know passphrase is sovable by a noob like me.
oh yeah, i'll be the first one to brag when i crack grany 3 ahaha
For some LLM problems(Witf), I got a flag with some prompt, but then when I re-run with the same prompt, it gives different output(not the flag). Will this be an issue later while submitting the NBs?
@olive ledge
As long as you achieved the flag once, there shouldn't be any issues. The LLM challenges are non-deterministic to some degree and we'll respect that.
Once you see it .... you see it ... its not the coding that is hard ... its seeing the solution 🙂
Is there a CLI command to submit for this competition? I accidentally submitted a file with empty(literally empty) flag, the submission gave an error, and the submit button on website is broken ever since. Cleared cookies, logged out logged in, changed browswer, none worked
create new submission notebook?
makes sense, let me try
"True. I probably woldn't talk about it if I were on top of the LB. "
not true. you can easily catch up
lots of "failed to process inputs" during the last 2h for semantle2
Appreciate your kind words, I may have more confidence if I manage to do simple counting for MNIST, lol...