PeerAiD: Improving Adversarial Distillation from a Specialized Peer Tutor in Neural Networks
Core Concepts
PeerAiD proposes a novel method in adversarial distillation, training a peer network to defend against student-generated adversarial examples, achieving higher robustness and natural accuracy.
Abstract
Adversarial robustness is crucial in security-critical domains.
Previous methods pretrain teachers but suffer from degraded robustness.
PeerAiD trains peer networks to defend student networks, surpassing pretrained models' robustness.
Results show significant improvements in robust accuracy and natural accuracy with various datasets and models.
PeerAiD
Stats
Adversarial examples are generated using Projected Gradient Descent (PGD) [30].
AutoAttack (AA) accuracy improved up to 1.66%p with ResNet-18 and TinyImageNet dataset.
Quotes
"PeerAiD achieves significantly higher robustness of the student network."
"We propose PeerAiD to make a peer network learn the adversarial examples of the student network."