Adaptive Perturbation for Adversarial Attack

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep learning models are increasingly vulnerable to adversarial examples.
• Current gradient-based attacks often use the sign function, which can alter gradient direction.

Purpose of the Study:

• To develop a more effective method for generating adversarial examples.
• To improve the attack success rate and transferability of adversarial attacks.

Main Methods:

• Proposing a novel gradient-based attack method that utilizes the exact gradient direction with a scaling factor.
• Introducing an adaptive scaling factor generator to optimize perturbations for individual images.
• Theoretically proving improved black-box transferability.

Main Results:

• The proposed method enhances adversarial example generation, achieving higher attack success rates with fewer perturbations.
• Experiments on CIFAR10 and ImageNet datasets demonstrate superior transferability compared to state-of-the-art methods.
• The adaptive scaling factor generator reduces computational overhead.

Conclusions:

• Utilizing exact gradient directions with adaptive scaling is more effective than the sign function for generating adversarial examples.
• This method offers a significant improvement in adversarial attack performance and transferability.
• The approach is compatible with existing gradient-based attack techniques.