3 ms·
I would be a bit concerned about adversarial attacks with this. I’m sure someone will be able to come up with an innocent looking PR that the system will always
by tehsauce 4y ago
I would be a bit concerned about adversarial attacks with this. I’m sure someone will be able to come up with an innocent looking PR that the system will always approve, but actually is malicious. Then any repo which auto-approves PRs with this could be vulnerable.
- videlov 4y agoThere are 3 categories of predictors that the model takes into account, here are some examples: (1) The code complexity, and its perceptual hash. (2) The author and their track record in the repository. (3) The author's past involvement in the specific files being modified. With that said, an adversarial from somebody within the team/organisation would be very difficult to detect.