This was the final project for CS182 at UC Berkeley. The goal was to build a computer vision system that is somewhat resistant to misclassification on natrual adversarial examples. This was particularly challenging for 2 reasons.
- This is an open problem in the field. No one, including top their reserchers know exactly how to do this.
- We were seriously limited by our resources. We had about 2 weeks including finals week and close to 0 compute power.
That being said we did manage to produce a training system that was able to outperform traditional models on adversarial examples from imagenet.
See out site here.