Towards Unified Robustness Against Both Backdoor and Adversarial Attacks

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Neural Networks

Background:

• Deep Neural Networks (DNNs) are susceptible to backdoor and adversarial attacks.
• These attacks are typically addressed as separate issues due to their distinct timing (training vs. inference).
• Existing defenses often tackle these vulnerabilities independently, leading to suboptimal protection.

Purpose of the Study:

• To investigate the relationship between backdoor and adversarial attacks in DNNs.
• To propose a unified defense strategy that addresses both attack types concurrently.
• To develop a method that leverages the connection between these attacks for enhanced model robustness.

Main Methods:

• The study establishes a link: backdoor implantation alters adversarial examples, and these examples resemble triggered images in infected models.
• A novel Progressive Unified Defense (PUD) algorithm is introduced.
• PUD employs a progressive model purification scheme, using adversarial examples to remove backdoors and then enhance adversarial robustness.

Main Results:

• The connection between backdoor and adversarial attacks is demonstrated to be ubiquitous across various backdoor attack types.
• The PUD algorithm effectively identifies poisoned images, allowing for imperfect initial datasets.
• PUD significantly outperforms state-of-the-art backdoor defenses and competes with advanced adversarial defense methods.

Conclusions:

• A significant, previously unrecognized connection exists between backdoor and adversarial attacks in DNNs.
• The Progressive Unified Defense (PUD) offers a unified and effective solution for defending against both attack types simultaneously.
• PUD represents a substantial advancement in securing DNNs against sophisticated threats.