ATR-UAP: Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Adaptive Truncated Ratio...

Area of Science:

• Computer Vision
• Machine Learning Security
• Deep Learning Robustness

Background:

• Universal adversarial attacks generate image-agnostic perturbations (UAPs) for CNN robustness evaluation.
• Existing universal attack methods often require large datasets, which are impractical in real-world scenarios.
• Current data-free universal attack methods lack adaptive layer selection, limiting their transferability to new models.

Purpose of the Study:

• To propose a novel data-free universal attack method, ATR-UAP, that overcomes the limitations of existing approaches.
• To enhance the transferability and generalization of data-free universal adversarial perturbations (UAPs).
• To improve the efficiency and effectiveness of robustness evaluation for CNN models.

Main Methods:

• Developed an Adaptive Truncated Ratio Maximization (ATR-UAP) approach for data-free UAP generation.
• Integrated adaptive layer-wise weight learning with exclusive sparsity and prior guidance.
• Introduced a variance reduction strategy to stabilize CNN activation values during training.
• Proposed a curriculum-guided alternating optimization algorithm to promote input diversity and address optimization challenges.

Main Results:

• ATR-UAP demonstrated superior performance compared to state-of-the-art data-free UAP methods across various datasets.
• The proposed method showed significantly improved cross-model and cross-task transferability.
• Evaluations confirmed enhanced generalization and robustness of the generated data-free UAPs.
• Effectively addressed the limitations of rigid layer selection in prior data-free methods.

Conclusions:

• ATR-UAP offers an effective solution for generating data-free universal adversarial perturbations (UAPs).
• The method significantly improves the transferability and robustness of adversarial attacks.
• ATR-UAP provides a valuable tool for evaluating and enhancing the security of CNN models in data-constrained environments.