FedCut: A Spectral Analysis Framework for Reliable Detection of Byzantine Colluders

Area of Science:

• Machine Learning
• Cybersecurity
• Graph Theory

Background:

• Federated learning (FL) is vulnerable to Byzantine attacks where colluding malicious participants upload harmful model updates.
• These attacks can severely degrade the performance and reliability of the global model in FL systems.

Purpose of the Study:

• To propose a general spectral analysis framework to detect and mitigate Byzantine attacks in federated learning.
• To enhance the robustness and accuracy of federated learning models against coordinated malicious actors.

Main Methods:

• Developed a spectral analysis framework to identify consistency and coherence in malicious model updates.
• Formulated Byzantine attacker detection as a community detection problem in weighted graphs.
• Utilized modified normalized graph cut and spectral heuristics for robust attacker identification.

Main Results:

• The proposed FedCut method demonstrates superior resilience against various Byzantine attacks.
• FedCut achieved 2.1% to 16.5% higher average model accuracy compared to state-of-the-art methods.
• FedCut showed 17.6% to 69.5% better worst-case model accuracy under attack scenarios.

Conclusions:

• FedCut offers a robust and effective solution for securing federated learning against Byzantine collusion.
• The spectral analysis approach provides a strong theoretical guarantee for convergence with bounded errors.
• Experimental results validate FedCut's significant performance improvements in challenging attack settings.