Improving Transferability of Universal Adversarial Perturbation With Feature Disruption

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are susceptible to universal adversarial perturbations (UAP).
• Existing UAP methods have limitations in transferability (data-dependent) or performance (data-independent).

Purpose of the Study:

• To propose a novel universal adversarial attack method for generating UAP with enhanced transferability.
• To address the limitations of current data-dependent and data-independent UAP approaches.

Main Methods:

• Devised an objective function to disrupt model-agnostic features, weakening significant and strengthening less significant channel-wise features.
• Utilized out-of-distribution (OOD) data for training UAP, eliminating dependency on labeled samples.
• Employed mini-batch input gradients and a momentum term for iterative UAP updates, capturing both local and global information.

Main Results:

• The proposed method demonstrates superior performance compared to existing UAP approaches.
• Achieved strong transferability of UAP across diverse models, datasets, and tasks.
• Effectively disrupted model-agnostic features for robust adversarial attacks.

Conclusions:

• The novel UAP generation method offers improved performance and transferability.
• This approach enhances the understanding and mitigation of adversarial attacks on DNNs.
• The findings contribute to developing more robust deep learning models.