Improving transferability of adversarial examples via statistical attribution-based attacks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks, which are crucial for assessing their robustness.
• Feature-level attacks corrupt intermediate features but often lack precise feature significance metrics and can limit transferability.
• Existing methods struggle with accurate feature importance assessment and optimization objectives in adversarial attacks.

Purpose of the Study:

• To introduce a novel Statistical Attribution-based Attack (SAA) method for enhancing adversarial attacks on DNNs.
• To improve feature importance representation and refine optimization objectives for stronger attack performance and robustness assessment.
• To address limitations in feature significance metrics and transferability associated with current feature-level attacks.

Main Methods:

• Developed the Statistical Attribution-based Attack (SAA) focusing on feature importance and refined optimization.
• Introduced Region-wise Feature Disturbance and Gradient Information Aggregation to compute Comprehensive Gradient for accurate feature representation.
• Employed a statistical attribution-based approach using average feature information across layers for an improved optimization objective.

Main Results:

• The SAA method demonstrated superior performance in adversarial attacks compared to existing techniques.
• SAA improved the attack success rate by 9.3% over the second-best method.
• When combined with input transformation, SAA achieved an average success rate of 79.2% against eight leading defense models.

Conclusions:

• The proposed SAA method effectively enhances adversarial attack performance by improving feature representation and optimization.
• SAA offers a more robust approach to uncovering DNN vulnerabilities and assessing model robustness.
• The method shows significant potential for evaluating and improving the security of deep learning models.