Universal Adversarial Attack on Attention and the Resulting Dataset DAmageNet.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks.
• Existing attacks require extensive knowledge of the victim DNN or rely on structural similarity and numerous queries.
• Improving adversarial attack transferability remains a significant challenge.

Purpose of the Study:

• To propose a novel adversarial attack method, Attack on Attention (AoA), that targets a common semantic property of DNNs.
• To enhance the transferability of adversarial attacks by modifying the loss function.
• To create a universal adversarial dataset, DAmageNet, for benchmarking DNN robustness.

Main Methods:

• Developed the Attack on Attention (AoA) method, which replaces the traditional cross-entropy loss with an attention loss.
• Integrated AoA with other transferability-enhancement techniques to achieve state-of-the-art performance.
• Generated 50,000 adversarial samples on the ImageNet validation set to form the DAmageNet dataset.

Main Results:

• The proposed AoA method significantly increases adversarial attack transferability.
• The DAmageNet dataset successfully defeated 13 well-trained DNNs, achieving error rates exceeding 85%.
• Even with defenses or adversarial training, models maintained error rates over 70% on DAmageNet.

Conclusions:

• AoA is an effective and easily integrable technique for enhancing adversarial attacks.
• DAmageNet serves as the first universal adversarial dataset, providing a valuable benchmark for DNN robustness.
• The findings highlight the need for more robust DNN architectures and defense mechanisms against advanced adversarial attacks.