#ai-village-capture-the-flag-defcon31
1 messages Β· Page 7 of 1
real nice
been working on granny for two days trying to match the models
and now i notice im not using pretrained weights

did last year's participants also have a discord channel to commiserate with one another?
yes
Even loading the weights i cant match the models π¦
Tried more than 30 at this points
seems like bruteforcing the api is gona be the way
really eager to see the solutions this year, to have a chance to actually try for a medal next year
tbh i cant even match the pytorch pretrained ones with the keras one π
"i cant even match the pytorch pretrained ones with the keras one" of course you can't. they are not the same one
in the first place, waht makes you think they are the same? second you can easily check if they are the same or not. lastly, google can esaily help
i dont know, i guess im not touching grass enough these days
tried keras, torch and kaggle models, no match π
has someone reached a perfect match ?
not sure it is possible, they also could do an extra train batch
looking throught the chat several people have claimed a match
in my case, not even remotly close tho haha
it shouldnt be a match right out of the box
but what worries me is do i have to try out all of the models to find out which one is on the api?
check the prompt
i did, up to 5 digits with some sorcery. turned out useless
let me rephrase
i know what model to use but not what weights
given that there are multiple pretrained ones
there are? i did not know that
assumed it to be the default and that it is the only one
keras and pytorch don't match after preprocessing ?
i didnt check that
i checked only the weights
is the preprocessing different between the two?
yep
Is pickle an LLM? That seems the only plausible thing considering no actual code executions. Surely the most frustrating problem
but none of them match with the API scores π
yea but thats not my concern
cant understand what im suposed to do on pixelated
OCR is working fine on whatever text i give it, just no clue what to give it
There was some interesting discussion already about preprocessing. I found the comments by hengck23 very interesting. But I have to say I also have not found the model yet. 
i know and im following them
whats bothering me is i dont know if every pretrained model (keras, pytorch imagenet v1, v2) will work if i find the right preprocessing
The host said somewhere in this chat, that the difference between the mobilenet implementations is not expected to contribute much.
yea that's getting hinty
thanks, it helps a lot that i don't need to implement this stuff in tensorflow
You're welcome. See this conversation for more details: #ai-village-capture-the-flag-defcon31 message
"The host said somewhere in this chat, that the difference between the mobilenet implementations is not expected to contribute much."
i don't recall reading that. which exact message by the host are you refering to?
image adversial attack are divided into black box and white box. you should google about them
See link above.
be careful. i would only say, there are methods that require for you to know the model (+ other processing) completely and methods that doesn't (complete no knowlege). there are in-between methods as well (i.e. partial knowledge). in short, google more to find out what methods you want to use and the necessary conditions to make them work
yeah, everyone is free to make their own interpretations of what host says
more knowledge is better of course
Insane ! Good luck on the last 2 π
Thx. I'll try, though I think the last 2 challenges are somewhat broken.
Wow, that's very impressive! @craggy beacon
any hint on granny 1 pls.
read this discord channel π
Which ones left for you @craggy beacon
Any resources for pixelated? Really cant understand where to go, i can get the OCR do detect any text, just no idea what to give it
ask chatgpt to brainstorm about it
I officially give up, silver is not that bad π₯²
Good luck for the tryharders of the last week!
I am happy with silver also, a gold will be better (I need a solo gold) ... it seems so close to achive yet so far ! I hope I have it to go through this another week !
Is solo gold required for Expert to Master ?
GMaster
what I have been telling myself for the past ~~two ~~ 3 weeks
Congrats @craggy beacon
Hard to keep up withπ
desperatly trying to get granny 1 and 2 for a measly bronze
alas, no luck
granny 1 is so slow ... It s such a pain !
Yes, anyone currently outside bronze will probably need to get to 20.
Only one more for me then. Might be doable π Semantle 2 at 0.97 and some ideas for pickle and pixelated.
Im 20 but im trying for silver
I got all the annoying ones left. Granny 1/2 are the most likely ones for me
Maybe inversion
Im not sure CIFAR is even worth trying from what ive seen in the chat
Yeah.. betting on granny 1/2 and pixelated
Congrats @craggy beacon on 25 - Very impressive! My guess is final 2 are granny3 and hush. But will be interested to find out after the competition.
I've got to ask because it's been killing me. @gusty warren When you made this comment was around the time when you solved passphrase. I now and realizing you might have been referring to something different all together. Am I reading too much into it?
No idea you were here, lol I've saw your YouTube video and it was great about the last year competition, thank you
Thanks, glad you liked the video! I've been lurking and posting be few memes π - I feel like my video from last year wasn't well organized but hoping to do one again this year and put more time organizing my thoughts beforehand.
If its not a secret, which 2 are you still working on?
Can you imagine if there are people with 25 or even someone with 26 waiting for the last day, last minutes to submit? Lol
its too risky strategy π
It is a valid strategy to keep it secret. If he tells what he solved the other top contestants can focus on this "solvable" task, instead of spreading their attention.
not valid now that we are close to the end. Someone that would have had 25 flags keeping it secret would just have lost the first place
I meant of course submitting but not telling in this chat what he solved...
ah sorry, yes keeping the information secret is indeed a good strategy at that point π
I mean he posted it for like 5 seconds but then wisely deleted the message π
No, continue please, with who will I complain about Inversion if you're not here.
all those people (edit: made a mistake and included some people from the club 23, mb)
hey @olive ledge or @limber flower, may you have a look at Granny 1 ? It seems it is down 80-90 % of the time. Thx
works ok now
not sure what changed
R.I.P 22.
if anyone fills in the last spot of prize zone, I give up
now i feel the OCR pain of the previous posters π€ π
why do you think prize zone is limited by 24? π
get 25 and you'll be 2nd
next 24 should indeed secure a prize, just because its unlikely that another 23 will get 2 flags in such a small amount of time I guess
I am surprised! Seems like WTF5 got fed up with our conversations and spat out the flag.
How can you be so sure
i am not sure, there is probably exceptions, but from what we can interpolate from the people and discussions here, appart from a few outliers that solved inversion or passphrase, most of the people with 22 flags are stuck on passphrase/inversion/hush/cifar/grany3.
- it was more a joke than something else π
ππ
I guess yeah
Lol, I litterally just subscribed to your YT channgel and then saw you here. What I meant is litterally what I wrote: the vocabulary/dictionary I used to search was likely too small so I ran out of good candidates to try. Fixed that and got the flag before long.
new hint just dropped 
The what
Passphrase has been solved!!??
TBH though, I may or maynot be misleading you all unintentionally. My solution is quite stragith forward and I believe many people should have tried the same thing. I am very surprised that many people still can't solve passphrase, I am starting to suspect that I did something different that I am not even aware of...
DJ Khaled: suffering from success
You might also have got all stars aligned at the right moment :p
Thanks. That at least confirms some of the things I've been trying. But passphrase has been really tough for me and I've gone down many dead ends. At least that confirms my suspicions about your previous comment.
exactly, it could be mere luck that I hit a lottery ticket
I wish i got it on inversion ahaha
imagine a lottery ticket in granny3 ....
i think for passcode a lot have been said also in the past that could mislead many people
granny3 i have a last idea that i could have tryied, but it implied mounting a cluster of server to query more efficiently :p
but it is not a dumb luck like lottery ticket you acted using your knowledge and with a bit of luck you got the flag
i think the luck is the thin layer that allows you to not wrongly deny a direction because of a dumb error/missing info on some kind of postprocessing
pickle is a very good example of it: you miss the sweat spot, you can end up into a hell of complexity
Luck is just when your cat jumps on keyboard and solve granny 3 unintentionally. And yeah, you just call it a day.
Does the static flag in inversion mean anything or is it just a random letters?
just rent A100, it's enough to exhaust all during lunch. what blocks me is what's to optimize
what i need is not cpu/gpu time, but different ips :p
ddos could cause ban LOL
when you run multiple kernels on different tasks (or on the same task with different seeds/approachs), you do exactly the same
well, it's a matter of scale
"what i need is not cpu/gpu time, but different ips :p "
you need to reduce your search space by working on the most probable pixels
yep I know, but I would have tryied another way that I think could be promising giving my current view of the problem but that I will not implement in any case because I am too lazy :p
hope granny3 is not the same as cifar, which can only solve by api call
? cifar is the exact problem that cannot be brutforce
Give me a little direction of the sweet. Otherwise sweating a lot
there is a new kaggle competition
so maybe there will new attack for next CTF: fool LLM text detector
It was not complex? Then what should I do with my 10000 lines of code to solve it, just kidding, I'm already lost there, but I'm dangerous π
has there been any clarification on passphrase? I managed to target specific scores, but I didn't get any flag.
there have been some hints in this chat
i havent even tried to solve it so i have no idea if theyre useful
take a look here
Thanks!, i'll take a look
IP is baffling at first, but you learn a bit and both fall pretty quickly together
you need to keep balance, not so dangerous, not so safe π
looks like the 5th spot for money has been taken. Good luck to the people targeting the 25 flags now!
there is still a week ... you never know ... for CIFAR and Passphrase its once you see it you get it .. so the flags for both are floating in the air ...
have you solved CIFAR?
congrats @severe pasture
no I didn't I tried in two diffrent domains (or directions) ... nthn so far
everytime I think I have something that fit ... its try again ... CIFAR could have logically different answers ... but there is only one answer will get the flag :/
"looks like the 5th spot for money has been taken. Good luck to the people targeting the 25 flags now!"
the doctor cannot save all patients.
we cannot solve all problems.
that 's the reality of life...
we can only try our best and there is always the next time
there is still a week , then break ( I hope) till the santa challange next month !
however the llm detecting ai generated text seems intersting ... no break π
is there any forum or website where you can find AI security competitions?
I really enjoyed this one, but I stumbled upon it on accident π
I can join you in complaining. been staring at ghostly messages all day. π
are there any doctors who could help with brain damage from CIFAR? π
maybe that's the way it should be solved XD
cifar is actually fun, guessing other's mind. like guessing other's ins password
There is last years competition! Checkout the discussion boards.
Thanks, I feel not alone now π Today, like yesterday, I don't if I should submit the message from the AI "as is" or if I need to pass TOEFL first
We'll all know in a week
I am probably not that far yet. I still struggle with 4,5,7 like previous posters.
Semantle2 done finally. Was stuck on 0.98 for a long time
you are not ... you just didn't see it
A week later when we see how CIFAR is solved ... we need to say "oh ... we didn't see that" ... its better than god I am stupid for wasting daaays
im also somewhat weirded out by this model matching
because i would have sworn that i already tried what turned out to work
i guess revisiting old ideas when youre well rested makes sense
I was able to match the model yesterday. Hopefully I can finish it tomorrow.
starting to think the problem is not the weights but the preprocessing zzz
im gonna try model matching aswell, cant get black box to work though there is probably some way around it
Haha, I've already gone through the solutions for last year too many times π
In pytorch or tf?
im not sure if im allowed to say hahaha
MNIST: 'Try again!' π‘
its so hard to shake off the idea I attempted because I feel like its the only thing that makes sense. Got myself stuck in some local mind optima
I believe in you. You can do it π
CIFAR 10 or CIFAR 100?
who knows π€·ββοΈ
"CIFAR 10 or CIFAR 100?" the answer is the 7 th part of the inversion static flag solution and 8th part of the hush static flag solution
sloth
Congratulation! So you also solved inversion?
I happy with sliver ! ... I ran out of ideas for the last four ... and I am exhausted :/
Feel you! 10 days of passphrase and still don't know what the objective is
hush needs a static flag?
Congrats! Did you get a high score for 4,5,7? or it is impossible?
Great! Still stuck here, I don't understand the way to generate static flag from what the AI is trying to tell
I would love to discuss, since I really enjoyed the problem, but i think it would be too much.
yep, solved it a few days ago
i see youve solved it as well, nice
What do you think is a good candidate to work on for 24? I havent yet tried CIFAR at all but what I read, it seems impossible.
passphrase probably. there are some recent hints in a chat, may help
either cifar or passphrase
i had way too many logical ideas for cifar but none seem to work
Is there some evidence for solved Granny3, Cifar or hush in this chat?
its more of you either see it or dont
cifar has been solved
moo mentioned it a while back #ai-village-capture-the-flag-defcon31 message
yesterday, or so, some 24 on the leaderboard was renamed granny3. But probably trolling or? π
thats what im hoping for
everytime I've changed my name it's been to troll lol
Looks like you figured it out π
Nah, passphrase is my nemesis. Solved another task π
Nice work!
0.9999921321868896, 'Granny Smith' please tell me I'm close
there are alot of clues in this channel, I suggest reading through the chat history
I cannot solve MNIST and CIFAR. I can't believe I can't even count.
Got 1.0 prediction for granny smith but no flag...
and I got the flag
finally
And granny2 is also solved. Finally I can enjoy my weekend haha
I miss old days .... what old days ? ... when flags used to show up π¦
17 days ago, as for me 
try rerunning one of the solved problems
i was checking my old notebook, my brain is fucked up, I get a dopanin shot when I see a output starting with {"flag":"gAAA.."}
are you guys going to other comps after this one ?
I'm already joined UBC-OCEAN, but not submitted any yet
yeah, i just got 18 flag in 2 days, its so sad now
gona start enefit next week
this is a marathon not a sprint
Was anyone else getting sporadic 504 timeouts for WTF? WTF3 in particular?
this is my sole opinion and maybe most would not agree. while the CTF tasks could be made more difficult, i prefer team work and also discussion. we all can go further+faster in a team and discover more security bugs . it seems that everyone is stuck.
but we have learned more that way because you can't delegate some stuff to others
how can I convert the image or tensor to base64 string for granny?
@ember relic Thank You!!

{"message": "An unexpected error occurred in scoring script. Check the logs for more info."}, I wish I can check logs
@olive ledge @limber flower
I'm trying to understand the connection between sentiment scores and the passphrase(password) concept but no luck so far
Lookingβ¦
Didnβt see anything particularly bad, service is still up. Check your inputs π
For those with 20+ {"flag":"gAAAAABl==GG-YourFire!"}
I went to the shower and I have an idea for passphrase, no joking π
Is this relevant to something you see in a shower environment? π
Inversion endpoint looks bad: "{'message': 'The server encountered an internal error and was unable to complete your request. Either the server is overloaded or there is an error in the application.'}
Inversion is unstable since friday to me.
Adding retry annotation around the query helps to workaround
Me trying to relate to all this bathroom stuff
but did your idea worked ? since a week I have idea strike but all fails π₯²
i had an idea for passphrase but its really hard to find a sentence with the same sentiment while keeping in mind the constraint
I'm still working on it.
Where's Grok @ with all its math might.
Why does hush sometimes returns 2 elements whereas sometimes 12?
Would love to know lol
still, I cannot count.
I wonder how simple it is.
Me too bro
looks like it worked π
Finally got Mnist...can't believe was so simple π€―
i atleast wanna know if the riddle really helps us find it or not
oh sweet semantle! finally
i'm at 0.97 for semantle2. I'm hoping I can get it before the comp ends.
@naive umbra check this out
Got the flag for semantle2!
so close and still so far in Inversion, I can't understand what is it, and also can't give up at the same time, strange feeling π
This hurts so much
Count Dracula is gentle and harmless compared to Count Mnist, and as for Count Cifar ...
So close yet cifar
Great prompt for a challenge for next year
just a quick question. i am also taking part in other kaggle competitions. but i find that their discord channels are pretty quiet.
am i right to say that that is the norm? and this discord channel is unusually active?
IT WAS THE FRIENDS WE MADE ALONG THE WAY
I guess part of the reason is that we cannot spam the discussion threads for medals for this comp, so we all came back to shitpost here instead
No π¦ still working on passphrase, I solved hush
Why is my shower not giving me any flag
I've learned to follow a path of continuous suffering in this competition.
just kidding... or maybe not π
I would have gone to the moon, if there were a hint of cifar there.
"Why is my shower not giving me any flag"
you took the shower at the wrong date
time, place, person
maybe we can build a model to find optimal shower parameters to get ideas
Keep oscillating between "not so dangerous" and "to much dangerous"...
shower thoughts confirmed!!!!!!
Wow! Congrats to 2nd place!
εοΌ
congrats
I think two with 25 is enough thoughπ π
because only one unsolved is granny3 pbly
Because of the existence of easter egg.
Im going for a walk, I better see a flag when im back
man the amount of walks ive taken in the past week
it doesnt work most of the time
Is this confirming you solved CIFAR? π§
imo the challenge creation process for granny3 was as follows:
||get an image, do stuff to lower accuracy, add this god forsaken pixel, lower accuracy more, remove the pixel||
and if that happens, ||imo, there is a solution, but it is somewhat degenerate in the sense that it would be out of place in the reference to the original, or after-second-lowering image structure, thus leaving it be in a part of search space that has no locality or continuity or something||
'cause otherwise the answer would've been found already by many people since the papers' algorithm ||uses locality and continuity that are coming from image structure, that may have been destroyed during the perturbations and|| may be useless here, requiring new approaches
it was 10k$ shower π congratz!
I think they just performed the 1 pixel attack on multiple images until they found one for which the pixel attack worked and provided a very different class from the original image
then they would not have been sure that the flag is obtainable
why not ? The pixel attack can be performed from the whole pipeline of the model, including the preprocessing
- it has been said by the organiser that a match of the model is not necessarly needed (they didnt use it to test the challenge)
then it would have worked for someone already
certainly the wolf image has been artificially altered for that exact model to give 28% acc because of the way that it is recognized by other models with 80+%
but I still do not know why this processing prevent differential evolution
For me its just that there is a lot of local minima, hence its very difficult to optimize.
meanwhile im not even able to solve granny1 with black box attacks π
I think I know why DE is not working well and I have a suggestion of algo to try for granny3 that I am now lazy to implement and that would take a lot of time and calls, i'll be happy to discute it once the competition ends
I have some thoughts on g3 as well but will wait for competition end to share, as should everybody π
there are alternate algos indeed yet imo DE should work best
that's the thing i was trying to say above - that global minima may be in an unfindably small region and have nothing promising say 1px to the side and 10 units of brightness lower/higher and no way DE or something like that will find that
Finally I got MNIST flag. I didn't expect this answer.
Same boat haha, no Blackbox and can't replicate model
another 25 π
its going to be a looooong three days ! π
How can you even figure out what the output of hush means. Crazy π
pickle is just so soul crushing... doesn't make sense to go for any other challenges since pickle is apparently so much easier than them, and here i sit having sunk 20+ hours into trying to be dangerous by sending pickles to a server where no code is executed... and no feedback is given. at least for mnist we know we have to count something. really not a fan of challenges where the goal is not given. rant over
Β―_(γ)_/Β―
vinegar + salt + sugar + cucumber + spices
stick in jar
But do you understand what is or isnβt dangerous?
Simply having a pickle file that executes code ISNβT necessarily dangerous.
For the same reason you canβt email PE files into anyone with even a smidge of security.
I solved pickle really by thinking how waving keyboards works...
And I still do not know why my solution is accepted.
Any of my attempts going even slightly in the direction of code execution get flagged as dangerous, and even some that don't, making me eliminate that as a potential means to being more dangerous, and then there is nothing left in hand. Anyway, I am not looking for hints, just needed to rant.
This is actually a great hint.
hint: wireless keyboards are a godsend
attach to pole and hang outside
boom
flag
being a complete newbie in AI security, Kaggle, or even coding, I can tell you thereβs a straightforward answer. In fact, thereβs a webpage that has a copiable answer. All you need is Google search
btw I have no idea on MNIST at all lol
my idea is to let chatgpt to read all messages in this discord discussion and kaggle forum
then we can make a kaggle CFT tutorial app
the important think is the "chain of thought" : how the winner thinks of this so0lution? and this is waht most people want to learn
we can ask winner to contribute their chain of though, solution code experiment results, what worka nd dose not ...etc . all content contributer for the app get gat paid roalty.
ultimately we want an app to give hint (not full solution) to guide us step by step
Prompt-Engineered. .
time to make AIV-7b a thing
The legendary shower should be encapsulated into an AI model
talking about gpt4, there has been an update, the current knowledge cutoff is now april 2023
is passphrase really about to having a sentence which return 0.3333 for each sentiment?
I can't understand what question asked for
"a difference sentence where everything is equivalent"
is it sentiment score? sentence meaning? none or both βοΈ
When you manage to get a different output you are generally on the right track in these types of challengers
Just got an idea for passphrase, but did not shower. So it's probably wrong. π
Omg, finally got passphrase. No shower needed
congrats
Congrats! What changed since this "ancient" message? The recent hints? π
Ahaha this message was on the 1st day of the comp so a lot changed. Regarding hints, I do not think qihuaz shared too much. What helped me is trying ideas and actually convincing myself that I had already invested too much time on this to give up - since passphrase is doable
idk how you can start solving anything before you solved cifar
im too pissed off to start anything else hahaha
Cifar is too frustrating. Twice a day I think I finally have the solution. Code it out for 10 minutes and only get "Try again!"...
I have a final idea to passphrase I going to try ...
Is English your native language? And does it help to understand what question asked for ?
i havent solved passphrase but just a heads up
in case you didnt see there was a different prompt for passphrase in the first few versions of the notebook
myb it helps
are you refering to the ctf-starter-aiv notebook?
i appreciate a lot the effort on developing this ctf. But I really don't see any value on these obscure tasks. It is better to have a hard objective than an easy objective that with unclear instructions, it is just frustrating π¦
yep
Just checked. Didnβt find any difference. All of them are βMy secret. β¦
I also cannot see any difference. Version comparison on kaggle does also not show differences in passphrase
wow what the hell
i have some kind of different prompt in mine locally
can you share the prompt?
They republished the notebook after some day, probably something wrong got there in the public notebook.
@olive ledge , is that correct?
Maybe the challenging part is understanding it. Whenever we understand what is the objective, doing part should not be that difficult
Collect flags by evading, poisoning, stealing, and fooling AI/ML
yeah, there was a change in variable name, but the prompt itself seems to be the same. Am I correct?
I think correct Answer to this questions would lead us to right direction π΅βπ«
Ok my latest and maybe final passphrase test is running ... let see the flag pleeeeaaaseee
Donβt over think it.
should we over think the other one π
Hallucination?
How much overthinking do you reckon is appropriate for cifar? π
any luck?
its going to take some time ... or luck ....
I only want one more flag, so i'm jumping erratically between all my unsolved problems and failing spectacularly at each π
same here 
I just want granny, ive put it way too many hours not to get it
even if it wont change my ranking
A sunk cost fallacy is a fallacy where someone convinces themselves that they should continue doing something because they put in a lot of time and/or effort into it, even if they actually get very little reward out of it. For example, a person might watch the first six episodes of Battlestar Galactica, but decides they do not like the show. How...
We all are victims of βsunk cost fallacyβ in CTFβ¦
And then there is βI havenβt failed 1000 times. The light bulb was an invention with 1000 steps.β
Any news ? π
No flag so far β¦ but I think I am on the right track .. but I need to speed things up , I am so sleepy so I will update the strategy on the morning ( unless a flag shows) hhh
This flag means gold so it,s better show it self ππ
hmmm.. people are still getting flags at the 20-threshold
i might fall out of bronze lmao
I won't believe it until I see it!
I'm trying to claw my way into bronze π
Stumbled upon this today: http://www.paulgraham.com/top.html π
"I realized recently that what one thinks about in the shower in the morning is more important than I'd thought."
I was able to load skops for cluster 1, but is there any tips on predicting with the data? It's probably simple but I'm not too familiar with skops
I did some basic searches on Google
(nevermind looks like it has the usual predict_proba and I just need to figure out how input it right)
My new idea failed me with passphrase ... π¦ .... there is one last edit I want to try ... then I am out !
Hi @olive ledge , Is granny3 fine? I see many 504 status code on my last optimizations
Goodluck, have fun.
I should have quit 12 days ago.
I tried resetting the cluster, let me know if it improves
Weird...Why the default code in pixelated challenge cannot run in my terminal, resulting in TypeError: Object of type bytes is not JSON serializable?
Sorry if that's a silly question. Does anyone have some clues?
Yes, it has been answered many times. Try adding .decode() after encoding with base64
append a .decode()
Many thanks. What should be done after adding this:
input_data = base64.b64encode(input_data).decode()?
import requests
import base64
with open('/kaggle/input/ai-village-capture-the-flag-defcon31/pixelated/reference.png', 'rb') as f:
input_data = f.read()
input_data = base64.b64encode(input_data).decode()
def query(input_data):
response = requests.post('http://pixelated.advml.com/score', data={'data': input_data})
return response.json()
query(input_data)
will return {'message': 'Invalid input.'}
data= => json=
Oh I see. Now Kaggle works normally
somehow the terminal run normally but outputs nothing = =
Thanks, that may be a bug in my computer
query(input_data) => print(query(input_data))
Is Hush slow for you guys too?? Its taking like a minute now to query
I have the same question. Mine is also far from 1
Granny1 seems a bit sluggish as well π¦
omg I still don't get what I need to do for Pixelated...
Neither I
Meanwhile I'm still wasting time on this one. It's the only one I still want to solve before the comp ends.
I need pixelated too. Wasted too much time on that
I know what text to enter but the OCR is so bad.
Any tips for Cluster 1? It's 3AM where im at, and im losing my mind trying to figure this out. π€£
knowing the text to enter is already a huge jump forward in my opinion...still far from that π¦
Dude I am trying it like from the past 4-5 days and still no progress with ocr π₯²
You can find some information regarding this in past messages
Idk if what I try to write is the right thing...but agree with you about OCR
haven't found anything useful here or anywhere else about pixelated...
use one same prompt solved all wtf1-6 except 4, looks like it can't break the securityπ
what do you mean? π
is chatgpt down ? I can't open my chat histories
https://status.openai.com/ looks good
Welcome to OpenAI's home for real-time and historical data on system performance.
score has to be 1.0 to get the flag right? I am at 0.018. soo far awayyyy π
Read the hint again
the one that says "find which subpopulation is being systematically misclassified" or the one that sounds like a harry potter line π€£
Which one is more clear to u?
@olive ledge semantle2 giving βfailed to process inputβ
ChatGPT is also down
Time to take a shower
Don't forget, this shower style will help you #ai-village-capture-the-flag-defcon31 message
use this instead of chatgpt:
https://chat.lmsys.org/?utm_medium=email&utm_source=gamma&utm_campaign=-lmsys-2023
and you can also vote which model response better
now my score is 0.0 and completely lost π
@olive ledge @limber flower The hush server is almost dead ... can you please check π
+1 for granny3, please
I will shower on 10th
Finally solved Cluster 1, after getting some coffee and burrito
We believe in you. Make sure you take a peek at the resourceβs discussions.
moo have you recieved the full printed mnist cifar dataset yet?
i spent extra to fedex πΊ
Yesterday I said my last try ... then I am out ... can anyone just tell my head that I am out ...
you are still in
I am not :/ ... I am just staring at passphrase .. hush ... and cifar ... granny3 no need to stare
granny3 is crop2 (from last year) to me
Are there any Cifar hints?
try not to go crazy while solving it
Are kaggle notebooks working for you? The code does not load for me π¦
same here... I've been unable to load any notebooks for the past 2 hs π
today nothing is working, kaggle, chatgpt, ....
Excelent timing 
My internet has been out all day
Appearently it may take a couple of days before its fixed π
somebody created really dangerous pickle
or gained admin privileges to gpt by uploading a image.
mischievous
wish I had one more week π
I figured out an approach to hush...
But it requires 660 petabytes of ram...
getting a new output from hush....anyone got this one at some point?
afaik wrong type of file or sample rate or codec or something like that
i had it sometimes, it was gone after matching properties of a file to send with the test noise
@olive ledge could you post the granny3 pixel right after the end of comp and leave the servers running for 3 days after? i reeeeeeally want to check if that needle in a haystack is findable and how close i got after wasting 2 weeks on it
anyone ever got this out of pixelated?
hope the thumbs up mean I'm in the right direction π
i am redoing granny 2 for a notebook I intend to put public.
So nice to see the probability going above 0.00069
I missed that feeling
when people still haven't realized that you can pixel shift the entire timber wolf with a checkerboard pattern and it will pass granny3 inspection
actually it would be really funny if granny 3 was actually not an advertial challenge, but a challenge to bypass the 1 pixel checker
25 hrs to be accurate
i feel your pain, inversion marked my game over here π’
two positions shy from silver medal
hmmm
Im so glad this will be over soon
I really want to see a black box solution for granny1
ive implemented 3 different papers without results
i must be doing something wrong
my black box solution for granny 1 work
There is one black box approach I havnt tried that I believe might work but I cba
if you want to experiment still, go on granny2 which is much less used, so has better latency
I got my mind on CIFAR instead, though its even less likely ill solve that one
I have tried LSB and the alpha channel but granny noticed the former and ignored the latter
I also checked if there is a color that granny may misrecognize because in the easter egg picture she wears a pair of sunglassesπ€£
Whaaat
I am getting a 0.94 score from Semantle 2. How (ir)realistic is it to complete the challenge by the competition deadline?
very
Very unrealistic?
very realistic
21 hours is plenty
time to shitpost for 21 hours straight π
Once the comp ends the torture ends as well hahaha
Theres some sleep to catch up on
Agree. Otherwise I just cannot stop wasting time on the big 4 but with no flag gained.
This is whatβll happen
any hints for Cluster 3? Few more hours left until competition is done, but my brain is fried because of Cluster 1 π
Right way. So it is a big hint.
i would suggest deleting your message
Sorry, I thought there were problems, I'm not so expert with CTF
I know this but no idea what to do next (to prevent accidentally giving hints I could not list my attempts though π¦ )
anyone? cluster 3? π€£
I am already π€£ . I know they are the same data set, my brain is just really fried
maybe need more coffee
for cluster3: try multiple times. There may be some reordering happening from time to time when you try to order the clusters
it took like 20-30 retries to get it
Hint: it has something to do with clustering
This is why i am serious in suggesting to play more with Cluster 2
I agree with Roy, with the right pre-processing, Cluster 3 become straight forward
i see. I did brute-force my way in Cluster 2 to get to the "number" π€£ , so i didn't do the clustering
π€£ soga
is something wrong with Kaggle? I can't choose GPU for my notebook, also adding data page freeze with no reason, is it me only?
avg time to solve pixelated in your opinion? ....I am really hating it btw ahah
I can't solve silly pickle. I'm probably going to bang my head against the keyboard when I see the solution
omg I am getting "please receipt for an offer" now...
5 min - 30 days
Thought I have figured out pixelated but turns out I was wrong π
I would be deeply thanks if people can share their brute force/genetic/whatever algorithm after this competition really. Get JsonDecodeError soooooooo many times...
is the prompt even useful for pixelated?...I am basically ignoring it now
pixaleted took me about 1hr to figure out the thing to do, and about 5 hours to fight against the OCR :p
Now I think I now what to write...but unable to make the OCR read it right way...
I have 14h now...
the key for pixaleted is to understand the output provided and what it means concretly. Cannot say more unfortunately
2 days to figure, 15-minute battle with OCR.
lucky you, I got stuck with the OCR because I was overcomplexifying my image at first :p
still stuck 22 flags btw ? didnt look at the ladder for some times now
Yes, dropped this competition 10 days ago and am currently participating in Google Fast and Slow.
I think I have the opposite problem right now π
First to solve goes 20...π
yes...but I have the feeling it won't be enough for a medal anyway....
Probably not enough
Goddamn this is amazing
I am fighting from the past 5 days sadly but still no signs of improvement in ocr
I will surely be kicking myself after viewing the solution tomorrow
I will kick myself if I find I made the solution but OCR failed to recognise my art lol
when you get close to the solution for pixaleted, you know you are getting close
I'm currently in bronze with 20 flags but also on the brink to be pushed out
But we are not yet there... currently only another one to get 20 will be 'temporarly' bronze
actual way to solve pickle:
Any last minute clue for pickle? π₯²π
google it
Pickle or waving keyboard?
maybe ask the quirky chef from Bitville how to make pickles?
Is anyone still with me in tackling passphrase? 
Me. But with no more ideas.
I gave up when I saw i could not get Inversion
with a cache containing 800k+ entriesπ₯²
actually it does, eventhough couldn't solve it yet
Not going to the million would be a shame
I fear that i have been unintentionally mislead. π I am deep down a rabbit hole and it seems strange that 2 people solved it after taking a shower within like 3 and the other 12 hours (maybe because he did hush in the meantime). π
Now we know why the sever is so slow π
I will give up...finally manage the OCR read...and found I was wrongπ¬
this competition was good to work on multithreading skills
I'm desperately submitting stuff for cifar, the api response now is "you done?"
a significant step from Try again
and multi-tasking π
if you dont cifar you need see the eye doctor.
Hey folks, new to Kaggle and CTFs in general. I already signed up for the competition before the deadline, and I'm just wondering how I will be able to see the answers when the competition is over.
with our writups
the winner is gonna write his notebook
Cluster 3 is destroying me right now π
Anyone else on semantle2? Stuck at 0.95.. Now I am going crazy
That's both letting the computer multithread and multithreading yourself!
Finally got pickle. I donβt know how but I got it.
Wasted only two weeks
Soul crushing
Tomorrow big pain for pixelated and inversion solutions...
and CIFAR because it is simple
Last 9hours!!
I need 4 showers
It is big pain now, no need to wait till tomorrow π¦
I'm almost confident that I have got first part of inversion, but somehow second part is not workingπ€
3x 23 and 1x 22 in the past hour from Russia, interesting
I am currently writing a notebook for Inversion which i plan to publish after the competition close. So you can have a look then. π
sounds familiar
out of curiousity, was there a lot of difference vs the inversion equivalent of last year ?
I really really hope I missed a big subtility in that flag
yeah me too, otherwise it will hit hard
too far away now. Just going to enjoy all the solutions as they are released. Make some good write-ups every1
not really
You need a thunderstorm
If I miss it because of a character missinterpretation, i'll be kind of pissed aha
Yes, also for it...I think I felt in overthinking
organisers posted some pretty good clues for inversion
where in this channel ?
Is it possible to have the full message from AI and do not get the flag?
I'm not gonna say anymore but the turning point for me was due to messages in this channel π€
I fell like I understood this one...but still unable to get the flag...and time is running fast...
What was your turning point for passphrase? π
the shower is a good turning point for that one
I printed it out all the clues from this channel on paper and highlighted the key parts lol
I already have a word document with screenshots. Probably need to print it out and laminate them so i can take them under the shower.
same
gosh i spent all the month here trolling and shitposting, I really doubt I miss a message, but if this is the case, i'd be even more pissed ahah
For me I had inversion from day one ... but couldn't see it ! .... I got it in the end ...but it was in a way like my problem with cluster3 ... wasted days on a silly mistake !
oh wait wrong reply
also looking at how long it took people to solve problems helps
you can narrow your search that way
yeah, for passphrase, definitly, qihuaz gave a lot
anokas was #1 for a long time, I think discord messages made a big difference
and yet its driving me crazy ! I had so much "it must be this or this" and nthn :((
there are good clues from very early on in this channel π
same π¦
@severe pasture you are not tryharding the last hours to go ITM ?
trying to get cifar
gave up on hush and my granny3 is still running in the background
I think I read them all, but now i have to go back again... π
you got a good score for granny 3 ?
nothing better than the 0.00069
but I think I know what the intended solution is, just wasn't able to crack it in time
i would have had a methodology also, but it would have taken too many calls to converge and I was lazy to implement it :p
just so exhausted at this point that I'm averaging like one cifar idea every couple hours by now
I bet the clue in the input data is not a clue like for mnist
i think what was missing for cifar is a clue actually part of the solution
it is a clue for mnist π
never figured it out lol
it is actually a clue for mnist
and I thought it would be helpfull for cifar... well we can discuss that later today :p
Thats a new unit for creativity: Cifar ideas/hour
rewriting the same 3 cifar ideas from scratch hoping there was a small bug in your previous tries 
same here for the past couple days
won't ever be able to look at cifar the same way again
chat is being hot today π
Just seen your profile, you will become Kaggle competition master tomorrow.
i hope so π unless 3x24 appear in last hour of comp. and it usally takes couple of days for medals, to drop cheaters and finalize lb
well, my strategy of doing granny3, using the solution to get granny1 and granny2 and then finishing up other stuff failed miserably
now my best hope is to get to bronze
If granny3 is to alter only one pixel, I do not think the result can also pass granny2 using compression
i will come so far as to say that the guy who got 0.000831 accidentally altered pixel on a local model after postprocessing instead of checking the solution on the server
true, but that would be just rerunning the algo for x pixels
which is trivial if the needle was found
but im 99% sure that granny3 is not what we all think it is.
it's like trying to drastically change chatgpt output by changing only 1 character in 32K tokens prompt
tbf that is actually possible
well changing to a specific token not a character
so next year challenge?:)
changing character can actually change token, so it's kinda possible. but it has to be deterministic model to make sense also
why not the Pile
simple... 
wrong shape - should be (36001, 20)
"Try again."
Try again!
we need to motivate people. "You're doing great. Try again". "It was wonderful idea, but try again"
I think the only result is to decreasing the sentimental score of 'great' etc
maybe atleast leave away the "simple" in the problem description, so we dont feel so bad. π
maybe it is simple
like granny 3, just pick the correct pixel. π
honestly, imo, for cifar mnist etc type of tasks there should be like 5-10 possible solutions responses to which guide to the exact solution ex. "Yes, you did X, Y and Z, but now dont do Z and do something else" or "Wrong order!"
or just straight up do 6-7 subtasks
or give us an interesting riddle like passphrase and hush
If I start now granny?
Remaining granny1-3, inversion, hush, cifar, passphrase
do Russians usually win the Kaggle contests? i also see a lot of Japanese winning
they did, like 3-4 years ago
you might solve granny 1-2
depends on the way youre gonna solve it
inversion took me like 2 hours
so id go with either granny1 or inversion
idk, I just thought that if there were 26 people with 23 in 30 days, three getting 23 within the same hour isn't something you'd expect
This happened earlier also π
Lets have faith on kaggle
must have bought some flags. I don't really care about winning, i just want to figure out Cluster 3 so i can have 4 points to my name π€£
why not do the easier challenges? π
i already did the easiest of them all, Test, Cluster 2 and 1. Just need Cluster 3, i got few more hours left. Been awake for 19 hrs. now
the difficulty is not ordered. some of the later ones are just chatting with an ai π
what is the flag 1-6 + spanglish + pirate are the easiest
yeah, the prompt ones are pretty easy
I have invested 99% of my brain cells in cluster 3, i can't get it out of my head π€£
IP ones too
look at last year's solution notebooks π
I tried to think diffferently of granny3 , I tried some stuff that if it worked would be the coolest ever (in my opinion :P)
I have spent days trying to solve inversion and I can't narrow down the candidates at all. Thanks, I will try granny.
Sloth, done.
As is probably obvious, difficulty depends on each person's skills and experience. I have 14 flags, but only solved two of the six WTFs (1 & 5). People experienced in hacking, prompt injection etc. apparently breezed through them. However, I found challenges like cluster1 & semantle fairly easy [needed time, but was always going to solve]. Some others, you just need to try the right thing - which could randomly be quick or slow. A few, you have to read the prompt carefully. And maybe one or two where you need to put yourself in the mind of the setter.
On earlier weeks where I was really active I got 1 answer straight up because someone gave a hugr hint for pixelate
Message deleted after 2 seconds but I had been stuck for days but got it within 5 minutes after seeing the hint
Also people openly talking about adverserial networks, black box attacks etc for granny
Even just saying that a flag was simple is a huge hint, not everyone is aiming for gold
I was very surprised that people shared that they were able to match the granny model (closely) locally
that is not something one should expect to be able to do for a random deep NN
I think a totally valid assumption would be that it is trained/finetuned on different data and there is no point in trying to match it
(we were given the architecture of course, but still)
I can only imagine what people were discussing behind closed doors considering how much was discussed in here. Ive gotten message from over 5 strangers in this discord asking for help and tips
I can only think of one way to identify cheaters and that is time. Because there is no other way as all solution are same( at least hard ones )
passphrase server prepare for one last attempt! π
I already gave up 10 days ago with 18 flags, and just waiting the comp to finish to show my work of trolling the llms 
nice π whats the timeline?
I have a new challange .... I am writing a notebook to share later and I want it to score in the end .... it scores 22 :/ !! .... what the ? How I should know which flag is wrong !
i encountered the same thing. I just repeatedly submited the csv file with a different line of flag changed to 'empty' and found it's wtf6.
if you want the notebook to be scored, you may need to hardcode the flags; the LLM-based challenges are non-deterministic so what may have worked at one point may not work later
I did a trick to slove this , I always have 23 flags ... but it scores 22
Found the problem , the pirate flag .... I fixed it
You will solve the WTFs if you just keep trying. Trust me they arent as difficult as you think
can we start talking about granny 3 ? I guess its too late anyway to implement any kind of ideas now ? :p
you can make 1k rps if you have local model, so better wait till the end of comp π
true true
@devout jasper
Cifar is simpler than mnist.
Are anybody solved mnist?
what π
cifar is simpler?? π
You solved CIFAR ?
Almost all task are kinda simple after you have solved them
i can tell y'all the exact 0.0006927 pixel
not like it matters at all
I think we all have it lol
i got it
the answer is clearly -12
NICE
Granny3 is simpler than Granny1.
Are anybody solved Granny1?
My dog just solved it, he has a PHD in woofness from MIT.
photo pls
I think i am there now... π now i just need to find the right dictionary and then i will claim victory in 40 min. π
I just got some imaginary numbers there, am I going right?
nah you are going up
This is my dog π he's a very good boy
I love this one
whhaaaat???
Imagine, for cifar the answer being an array of 0 
add π if you just tried array of 0 for cifar after this message
I already tried that days ago 
Same and it's probably not the dumbest thing I've tried
I am not sure if I prefer this to be trolling or not
both options are great π
Do you think they will just release the solutions for the unsolved problems or give hints like last year? I would prefer just hints.
How do I get credit in cluster3. Not able to rechar
I think moohax mentioned somewhere in the chat that they would release stats and hints after the end of the competition
Imagine another input_data being given as clue for cifar
Or more of those riddles π
im 100% sure i thought of the solution to cifar but never implemented it correctly
how is it possible to get "Invalid length" from passphrase when the length of the string is 40?
I think LLM models process input as list of words, not chars
waiting for the last minute cifar hint
wth, inversion in 2 hours? π₯² is it this much simple, I'm strugling with it not over a week
i was lucky that i found the right approach fairly quickly
damn that granny easter egg is fire
also i've got granny 1 and 2 just in time with a single image!
i dare to say that i secured bronze
man, if only granny3 was more solvable and i did not waste millennia cracking it
That's a hard earned bronze
that's one hell of a journey yea
first actual competition on kaggle also
except the university community one a while back but that did not count towards medals and such
although, granny1-2 woud have taken waaaaaaaaaaay longer without whitening the box which i did while trying the granny3
Stuck on 0.99999 π
Damn was stuck with this blackbox approach and trying to grind to 1.
Thanks
and also don't send apple image π
furthermore, there were screenshots here of someone actually making it to 1.0
and others like 4e-9
the whole point of these type of competitions is for people to find out things by themselves... you are sharing too much (again)
is someone preparing a notebook for hush and passphrase? cant wait to read the solution and upvote. π
I have a write-up ready. No hush/Cifar though
you got cifar? πͺ
Nope, I meant I did not get hush & Cifar
so, no granny 3 and cifar on anyone?
according to hosts there are people with cifar
Yuri got cifar ? I missed the info
i believe it's the other way around
no moohax said plenty of people got cifar
i will post hush but didnt have time to comment it correctly
i think it's more likely to get hush than cifarbecause of feedback that the server gives
he never said
I saw the message of yuri and it was not cifar π
i mean he did, but then deleted the msg supposedly
aint no way more people got cifar than hush
granny3 wasnt solved?
not that we know of
no i believe not
omg I got granny 3
no way
damn
time to writeup, don't worry...btw it was great chatting with you guys
may i post my solutions here after the comp is over? i'm too lazy to write a complete writeup
same, it's been great talking to you here!
it's a pity that not all competitions have this much activity on discord
probably the nature of the competition itself plays a huge role...different challenges compared to one big task like others
congrat to you all and to the winner (not going to say the name because I'm italian and you know, we are superstitious)
thank you @olive ledge and the other hosts once again
well, time to say goodbye to all that
Laziest solution is just make the notebook public
all the arxiv papers for granny3? 
which one of the 20 notebooks?
All?
You will not believe my granny score π€£
Did you plan on giving out hints for unsolved problems after the competition end and if yes then when? π
and all the font sites for pixleated that i could not solve
sadge
and like 10 github repos for various sound and image processing
it was fun tbh
but damn my ram had some work to do
I feel like I could do a really lazy post about pixelated, because the OCR worked without issues for me, main issue was my own editing skills, not the fault of the model
just show which kind of images worked easily for me
alright going to sleep 2am in my timezone bye all and good luck to the last hour grinders
please do
someone else might scoop me, but if noone does a better one, I will π
Is .... apple, a hint to cifar :)))?
No hints, ever.
So granny3 will remain unsolved? π
Will you reuse challenges or keep them as a secret for eternity?
We're exploring the possibility of having some permanent challenge infrastructure.
In that case no hints are fine. I just want to see a solution before the servers go offline π
π
Guys, just to say such CTF competition is really fun, hope so see the same next year! Bye.
So, some that remain unsolved may get a growing bounty and stay up indefinitely? Sounds really cool if that were to happen
Still up? π
imagine now someone post a sub with the 27 flags
It's very hard to go, it's like going to sleep at 23:00 at Christmas Eve
will the delete accout be excluded from the leaderboard?
I need 21 execution please to go to gold
ahahaa I need like 3 to go silver
I need like 80
Normally it takes a couple of days to finalize LB, Kaggle team will probably remove the deleted account then.
I hope the ip llm knows your ips
silver is top 5% no ? So theoretically, to push you to the silver zone you could also ask 50 people to join the comp :p
Need more time...
Almost no changes to LB within the last couple of days and so much progress within the last hours...
its like during an exam
I learned a lot in this challenge, I like how challenges were in diverge direction, now I'm in silver zone and I'm quit happy with result π
thanks for hosting, and thanks for this great channel and hot discussions
hope to see you all again next year π (also in other competitions)
@glass bay at the end you solved granny with a white box attack ?
Whitebox. My local copy turned out to be close enough
Mathematical maturity is a key concept for the professional life of a mathematician. The term "mathematical maturity" is often used by college and university mathematicians. As I have discovered in the last few weeks at PCMI, it is not a common term among K-12 teachers, though few have any problem in recognizing what it basically means. This i...
Provided to YouTube by Portrait/Epic/Legacy
The Final Countdown Β· Europe
The Final Countdown (Expanded Edition)
β 1986 Epic Records, a division of Sony Music Entertainment
Released on: 1988-09-16
Vocal, Composer, Lyricist: Joey Tempest
Bass Guitar: John Leven
Drums: Ian Haugland
Keyboards: Mic Michaeli
Guitar: John Norum
Mixing Engineer, ...
there's a hint for cifar in here somewhere π€
the hint for cifar was given if you kindly ask to the AI god
I prepared an horrible notebook with my results. Just 27 minutes before making it public. (I'm at 22 flags, still trying to go to 23, but don't expect granny3,inversione,cifar,hsuh or passphrase solutions in the notebook).
For the other challenges you should find something in my notebook.
I want to share and learn from others so feel free to criticize me and my work a lot!
Anyway it's like I'll share it and then I'll go to sleep 0:37 AM here.
Tomorrow I have to wake up to work π₯²
ask GPT gods, or even LLM gods in parituclar
the ai god was the agent of CTF5 I think π
I did the same challanges as you, 95% done with inversion though but don't have time to put all the pieces together. The remaining 5% could take 95% of time π
for granny3, did some of you tryied to implement the horrible TLMFO algo ? :p
considered it but never tried, did you?
yeah, cryied blood tears to make it worked... It did converge pretty fast to the usual pixel, but got stuck there as well very fast
i'll not share that code, not proud about it π
You can implement algos from paaers?
I just throw bunch of gradient descent around and hope it works π€£
i have like 10-20 notebook versions with different approaches I tried for granny3.
In the next 20 years maybe I'll find the time to sort them out and publish once everything is in place.
i had 20 notebooks trying to do the same thing with different seeds :p
That is also a way to go.
Multithread. Just at notebook level
need to mix kaggle, colab and local to avoid rate limiting too lol
just buy some proxies π
multithread + if response.status_code != 200: time.sleep(120)
I seriously considered mounting a cluster of t3 micro for my ultimate try for granny 3
Just buy enough VMs to test on each a single pixel change locally.
Downside: requires massive budget, I think way more than the money in the current world economy
You think the whitebox would not work for granny3?
i went for blackbox for all 3 tasks personnaly
same, never managed to recreate the model locally
i think i tried for 10min to replicate the model and then said to myself "naaahh"
I tried bluebox yesterday.
Promising...
squared attack worked very well for g1 & 2
Did some math and with some optimizations iirc it could be done in ~1300 hours on my machine haha
Next year I think we should have the famous sandwitch function already there in the notebook.
"""
import time
def sandwitch_decorator(f):
def sandwitch_f(input_data,sleep=0):
try:
time.sleep(0.1)
result=f(input_data)
except:
time.sleep(sleep+1)
result=sandwitch_f(input_data,sleep+1)#make a sandwitch as requested on discord
if "flag" in result:
print(result["flag"])
return result
return sandwitch_f
"""
I literally made a sandwitch decorator. So that I can just add it before the query method whatever it does inside.
that's definitely not the intended solution, I hope so at least
Not the correct sandwitch for sure. But still a sandwitch it is
for g3, my last idea (never tried), was to do a double DE algorithm
keep sharing info so I can get my silver because you'll be disqualified
please
π
if someone can solve granny from scratch in under 10 minutes they deserve gold
π€£ At least for the momnt I think there are only granny 3 congectures, so no disqualifications. But yeah, if 40 people get disqualified I can be closer to gold, so go on.
a master DE finding the (x,y) with a fitness score based on the optimal RGB for that pixel (found by DE)
good luck in implementing this in 8 min and having a converging solution :p
Initially, I didn't consider making different notebooks lol
Just realized how big that one ended up
how many api calls did you need for granny1/2? I think it was ~6k for me but my algo was quite inefficient
around 7k for me
Don't you have a quantum computer that can just use its "oracle" to solve it by extracting its resulting quantum state?
4
You just need enough qbits
Might as well brute force and do de for each x y
5 mins and I'll cry looking at inversion/cifar solutions
I did that for some whitebox gathered suspects, no luck
4 minutes
yeah, more extrem, but 786x786 optims would have been slightly too long :p
Ye like 12 days or smth
but I think the double DE could have been duable with could multi threading in a few days
moohax is tipying a lot.
I think he is going to receive the printed cifar dataset in some minutes
And he will... See... Far...
did some people used a DT for Cluster 1 ?
this haunted me also during the beginning of the comp
for granny3 I have a suspicion that there's a way to sneak multiple changes passed the API, like having it only check for changes after resizing the 768 image to the model's expected size but was never able to get it to work
2 minutes before understanding the meaning of life , how to see far and what is the secret sauce.
I aways hear Linkin Park - In the End in my head when i think about CIFAR: I tried so hard and got CIFAR. But in the end it doesn't even matter. π΅
same
1 minute
did you notice there was a cropping of the original image ?
maybe it had something to do with the easter egg
I tried sneaking 2 pixels
wdym?
its over !
I SURVIVED
No, even if they are cropped, they still are checked
even if you change 2px in cropped part, it failed
are we allowed share solutions now?
π₯²
can we post our solutions now?
We all survived
Wooooooo comp over we can share, right ?
omg
I can't believe it
yeah but i was wondering if the people using the white box approach saw the cropping
congrats @craggy beacon
please send cifar solution π
its time for the cifar solution
Alright time to drop the multi pixel granny attack
inversion is "letmeout" btw π€¦ββοΈ
did someone solve grann1 with gradient estimation, no substitution model?
Thanks