Real-World Adversarial Defense Against Patch Attacks Based on Diffusion Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are vulnerable to adversarial patch attacks, compromising their real-world robustness.
• Developing effective defenses against these attacks is critical for reliable AI applications.

Purpose of the Study:

• Introduce DIFFender, a novel diffusion-based defense framework against adversarial patch attacks.
• Leverage text-guided diffusion models and the Adversarial Anomaly Perception (AAP) phenomenon for patch detection and restoration.

Main Methods:

• Utilize a unified diffusion model framework integrating patch localization and restoration.
• Employ efficient few-shot prompt-tuning for adapting pre-trained diffusion models.
• Demonstrate adaptability to both visible and infrared domains.

Main Results:

• DIFFender accurately detects and locates adversarial patches by analyzing distributional anomalies (AAP).
• The framework shows robust performance in image classification and face recognition tasks against adversarial attacks.
• Evaluations confirm versatility across different settings, classifiers, and attack types.

Conclusions:

• DIFFender offers a significant advancement in adversarial patch defense strategies.
• The framework demonstrates flexibility and generalizability, defending against both visible and infrared adversarial patch attacks.
• A universal defense framework is proposed, adaptable to various domains and attack vectors.