Exploring Robust Features for Improving Adversarial Robustness.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are powerful but vulnerable to adversarial attacks.
• This fragility limits their use in safety-critical systems.
• Adversarial examples (AEs) are inputs designed to fool DNNs.

Purpose of the Study:

• To improve the adversarial robustness of DNNs.
• To identify and utilize features invariant to adversarial perturbations.
• To enable efficient adversarial example detection.

Main Methods:

• Proposed a feature disentanglement model to separate robust, non-robust, and domain-specific features.
• Trained a domain discriminator to distinguish between clean images and AEs.
• Conducted experiments on five diverse datasets against various attacks.

Main Results:

• The proposed model significantly enhances adversarial robustness compared to state-of-the-art methods.
• The domain discriminator achieved near-perfect accuracy in identifying domain-specific features in AEs.
• Adversarial example detection was achieved without additional computational overhead.

Conclusions:

• Robust features extracted by the model improve DNN defense against adversarial attacks.
• The feature disentanglement approach facilitates effective and efficient AE detection.
• This method preserves clean image accuracy while enhancing robustness.