Red alarm: Controllable backdoor attack in continual learning.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Learning

Background:

• Continual learning (CL) trains models on sequential tasks, facing catastrophic forgetting.
• Security vulnerabilities, particularly backdoor attacks, are under-explored in CL.
• Existing CL research primarily focuses on mitigating performance degradation, not security threats.

Purpose of the Study:

• To investigate the security challenges posed by backdoor attacks in continual learning.
• To define a threat model for attackers in the CL setting.
• To propose a novel, controllable backdoor attack mechanism for CL.

Main Methods:

• Development of a threat model for backdoor attacks in continual learning.
• Design of a controllable backdoor attack mechanism named CBACL.
• Empirical evaluation of CBACL on benchmark datasets like Split CIFAR and Tiny Imagenet.

Main Results:

• The proposed CBACL mechanism demonstrates effectiveness in launching controllable backdoor attacks within the CL paradigm.
• Experimental results validate the feasibility and advantages of the CBACL approach.
• The study provides insights into attacker capabilities and challenges in CL security.

Conclusions:

• Backdoor attacks represent a significant security threat in continual learning environments.
• The developed CBACL mechanism offers a new tool for understanding and potentially defending against CL-specific attacks.
• Further research is needed to develop robust defenses against such sophisticated security threats in CL.