Toward Understanding and Boosting Adversarial Transferability From a Distribution Perspective.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning Security

Background:

• Transferable adversarial attacks pose a significant threat to deep neural networks (DNNs).
• Existing research primarily focuses on model-centric explanations for attack transferability.
• The precise mechanisms driving adversarial transferability remain incompletely understood.

Purpose of the Study:

• Investigate adversarial transferability from a data distribution perspective.
• Propose a novel method to enhance attack transferability by manipulating image data distributions.
• Improve the effectiveness of both untargeted and targeted adversarial attacks against DNNs.

Main Methods:

• Hypothesize that deviating images from their original distributions enhances adversarial transferability.
• Develop a method to craft adversarial examples by manipulating image distributions.
• Conduct extensive transferable attacks on multiple DNNs to validate the proposed approach.

Main Results:

• The proposed method significantly boosts the transferability of adversarial attacks.
• Achieved state-of-the-art performance in both untargeted and targeted attack scenarios.
• Outperformed previous best methods by up to 40% in certain cases.

Conclusions:

• Data distribution manipulation offers a novel perspective on understanding adversarial transferability.
• The proposed method provides a powerful tool for generating effective transferable attacks.
• This research offers valuable insights for developing robust adversarial defense strategies.