Edge-Cloud Collaborative Defense against Backdoor Attacks in Federated Learning.

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Distributed Systems

Background:

• Federated learning (FL) in edge computing enables collaborative training for IoT intelligent services but is vulnerable to backdoor attacks.
• Backdoor attacks on edge nodes can rapidly compromise the entire network, posing significant risks to security-sensitive applications.
• Traditional defenses struggle with edge device limitations like bandwidth and unstable networks, hindering model retraining and global updates.

Purpose of the Study:

• To propose a novel layered defense framework to address backdoor attacks in edge-computing intelligent services within FL.
• To enhance the security and reliability of edge AI by detecting and mitigating malicious data at both edge and server levels.
• To maintain high model performance on the main task while defending against sophisticated attacks.

Main Methods:

• At the edge: Employing a gradient rising strategy and attention self-distillation to maximize data-category correlation and train a clean model.
• On the server: Implementing a two-layer backdoor detection mechanism to filter malicious updates.
• Server-side model restoration using the attention self-distillation mechanism to recover performance post-defense.

Main Results:

• The proposed two-stage defense framework effectively weakens backdoor attack effectiveness at the edge and server.
• The defense mechanism successfully eliminates backdoor updates, enhancing overall model security.
• The final model precision on the main task is comparable to that of a clean, undefended model.

Conclusions:

• The layered defense framework is well-suited for securing edge-computing intelligent services against federated learning backdoor attacks.
• Combining edge-level mitigation with server-level detection and restoration provides robust protection.
• The approach ensures model integrity and performance in challenging edge environments.