DPC: Dynamic purification chain for adaptive adversarial defense.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are crucial but vulnerable to adversarial attacks.
• Adversarial purification methods show promise but can be bypassed by strong adversaries.
• Existing defenses struggle against adaptive attacks that exploit defense mechanisms.

Purpose of the Study:

• To develop a robust defense against strong adaptive adversarial attacks.
• To enhance the security and reliability of deep learning models.
• To introduce a novel adversarial purification technique that overcomes current limitations.

Main Methods:

• Proposed the Dynamic Purification Chain (DPC) for adversarial defense.
• Combined pixel and geometric transformations for perturbation elimination.
• Utilized a feedback algorithm for dynamic chain construction to prevent over-purification.

Main Results:

• Demonstrated superior performance of DPC against strong adaptive attacks.
• Maintained high accuracy on clean data, indicating minimal performance degradation.
• Experiments conducted on CIFAR-10, CIFAR-100, and Imagenette datasets.

Conclusions:

• DPC offers a promising solution for defending deep learning models against sophisticated adversarial attacks.
• The dynamic and adaptive nature of DPC enhances its resilience.
• The method provides a balance between robust defense and maintaining model utility.