Robust Perception through Analysis by Synthesis
The intriguing susceptibility of deep neural networks to minimal input perturbations suggests that the gap between human and machine perception is still large. We here argue that despite much effort, even on MNIST the most successful defenses are still far away from the robustness of human perception. We here reconsider MNIST and establish a novel defense that is inspired by the abundant feedback connections present in the human visual cortex. We suggest that this feedback plays a role in estimating the likelihood of a sensory stimulus with respect to the hidden causes inferred by the cortex and allow the brain to mute distracting patterns. We implement this analysis by synthesis idea in the form of a discriminatively fine-tuned Bayesian classifier using a set of class-conditional variational autoe ncoders (VAEs). To evaluate model robustness we go to great length to find maximally effective adversarial attacks, including decision-based, score-based and gradient-based attacks. The results suggest that this ansatz yields state-of-the-art robustness on MNIST against L0, L2 and L infinity perturbations and we demonstrate that most adversarial examples are strongly perturbed towards the perceptual boundary between the original and the adversarial class.
READ FULL TEXT