Reproduce “Cross-Task Attack: A Self-Supervision Generative Framework Based on Attention Shift” on OpenResearch

About this paper

Qingyuan Zeng, Yunpeng Gong, Min Jiang

Researchers from Xiamen University developed Cross-Task Attack (CTA), a self-supervised generative framework that creates adversarial perturbations by shifting the collective attention of diverse AI models from important to neglected regions. This method reduces accuracy, mAP, and mIoU across classification, object detection, and semantic segmentation tasks, even against adversarially trained models.

Key insights

• Adversarial attacks can be effectively generalized across diverse AI tasks by manipulating the common attention mechanisms of the underlying neural networks.
• A self-supervised approach, leveraging 'co-attention' and 'anti-attention' derived from existing models, can overcome the practical challenges of obtaining multi-task ground truth labels and harmonizing varied loss functions.
• Shifting model attention to unimportant regions provides a robust attack vector that remains potent even against adversarially trained defense models, indicating a fundamental vulnerability in AI systems.

Summary by alphaXiv.