Attention distraction with gradient sharpening for multi-task adversarial attack.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Deep Learning

Background:

• Deep neural networks (DNNs) excel at visual tasks but are vulnerable to adversarial examples.
• Existing adversarial attacks primarily focus on single-task scenarios, limiting their effectiveness in multi-task learning systems.
• Multi-task systems require adversarial examples that leverage shared knowledge and minimize task-specific interference.

Purpose of the Study:

• To develop a multi-task adversarial attack method for deep learning models.
• To generate adversarial examples that generalize effectively across multiple tasks.
• To enhance the attack performance on multi-task learning networks.

Main Methods:

• Proposes a multi-task adversarial attack leveraging attention distraction and gradient sharpening.
• Attacks attention heat maps to exploit generalization information.
• Employs gradient sharpening to prioritize multi-task information over task-specific details.

Main Results:

• The proposed method improves the generalization ability of adversarial examples across multiple tasks.
• Demonstrates superior attack performance compared to single-task adversarial attacks on multi-task networks.
• Experimental validation conducted on NYUD-V2 and PASCAL datasets.

Conclusions:

• The novel multi-task adversarial attack effectively enhances adversarial example generalization.
• This approach offers a more robust defense strategy for AI systems handling multiple tasks simultaneously.
• Future work may explore further optimizations for multi-task adversarial robustness.