Enhancing Adversarial Transferability With Cost-Efficient Landscape Flattening.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning Security

Background:

• Adversarial examples demonstrate vulnerabilities in machine learning models.
• Transferability of adversarial examples, especially targeted ones, is crucial for practical attacks.
• Existing methods focus on optimizing perturbations at loss landscape maxima, ignoring minima.

Purpose of the Study:

• To theoretically and empirically demonstrate that flattening the input loss landscape enhances adversarial transferability.
• To propose a novel, cost-efficient attack method that considers both local maxima and minima for improved transferability.

Main Methods:

• Theoretically analyzed the impact of loss landscape flattening on adversarial transferability.
• Proposed the Cost-efficient LandscapE Flattening (CLEF) attack.
• Utilized gradient reuse for optimizing towards local maxima.
• Employed probabilistic modeling for optimizing towards local minima, enabling pre-training.

Main Results:

• Demonstrated that optimizing perturbations at both local maxima and minima flattens the loss landscape.
• The CLEF attack significantly improves targeted adversarial transferability compared to existing methods.
• Probabilistic modeling for minima optimization can be pre-trained, reducing attack cost.

Conclusions:

• Flattening the loss landscape by considering both maxima and minima is an effective strategy for enhancing adversarial transferability.
• The proposed CLEF attack offers a cost-efficient and powerful method for generating transferable adversarial examples.
• This research opens new avenues for understanding and improving adversarial robustness and attack methodologies.