ZeroPur: Succinct Training-Free Adversarial Purification
The episode discusses "ZeroPur: Succinct Training-Free Adversarial Purification," a method that defends models against adversarial attacks without retraining or training auxiliary networks. The hosts detail its two-step process, Guided Shift and Adaptive Projection, explaining how it uses a simple blur to push images back onto the natural image manifold. The discussion concludes by highlighting its practical benefits in deployment.