Improving Adversarial Robustness via Attention and Adversarial Logit Pairing.

Area of Science:

• Computer Vision
• Machine Learning Security

Background:

• Deep neural networks excel in visual classification but are susceptible to adversarial attacks.
• Adversarial examples pose a significant threat to the reliability of AI systems.

Purpose of the Study:

• To develop an improved defense technique against adversarial examples in deep neural networks.
• To enhance model robustness and accuracy when faced with adversarial perturbations.

Main Methods:

• Proposed Attention and Adversarial Logit Pairing (AT + ALP) technique.
• AT + ALP encourages similarity between attention maps and logits of clean and adversarial example pairs.
• Evaluated AT + ALP against strong Projected Gradient Descent (PGD) attacks with varying perturbation levels and iterations.

Main Results:

• AT + ALP demonstrated superior accuracy on adversarial examples compared to adversarial training.
• The method effectively increased average activations in key areas for adversarial examples.
• Achieved state-of-the-art defense performance on diverse datasets, including the 17 Flower Category Database.

Conclusions:

• AT + ALP significantly improves the robustness of deep neural networks against sophisticated adversarial attacks.
• The technique focuses on discriminative features, enhancing model resilience.
• The proposed method represents a substantial advancement in adversarial defense strategies.