josephtlucas/hackthisai

Adversarial Machine Learning (AML) Capture the Flag (CTF)

hackthisai's Issues

Challenge: Evade airport screening

Use https://paperswithcode.com/dataset/hixray in an evasion or poisoning scenario

Challenge: Steal this stock market model

With either a classification [buy, hold, sell] or regression [prediction] model for stock data, expose an API to players and challenge them to steal the model.

Optional increase in difficulty: Rate limit the API

ModuleNotFoundError in hard_phoebes_smorgasbord

So far it has been great fun working through your challenges. Thank you for setting them up!

Unfortunately, I have an issue running the hard_phoebes_smorgasbord container. The line from pytorch_tabular import TabularModel raises the following error: ModuleNotFoundError: No module named 'torchtext.legacy'. It seems that the latest version of the libraries is incompatible.

After downgrading to torch==1.7.1 and torchtext==0.8.0, the error went away. But I am worried whether this really solved the issue, because the predict endpoint gives me nan for the example.csv. Is this the desired behavior? (I didn't want to look in the challenge code so as not to spoil the fun.) Would you mind sharing which library versions you used while developing this challenge?

Challenge: Hide the celebrity

With a model trained to identify a celebrity, try and "hide" (misclassify) the celebrity by changing a minimal number of pixels (set boundary for acceptable submission).

Challenge: Evade malware classifier

Test some set of required malicious functionality

Challenge: A/B Testing

For a hard challenge: steal 1 model from an A/B test

Flask-ify

Convert existing challenge submissions to Flask API endpoints to streamline play flow, increase realism, and expand technologies players can use.