Enhanced Spectral Clustering Robust Aggregation for Lens Detection in Federated Learning Against Byzantine Attacks

Area of Science:

• Machine Learning
• Cybersecurity
• Distributed Systems

Background:

• Federated learning (FL) enables decentralized training but is vulnerable to malicious clients introducing inaccurate parameters.
• Existing defenses struggle against sophisticated, gradually evolving Byzantine attacks.
• Accurate identification of malicious updates is crucial for maintaining model integrity in FL.

Purpose of the Study:

• To propose an unsupervised approach for identifying malicious updates in federated learning.
• To enhance the robustness of federated learning models against Byzantine attacks.
• To develop a defense mechanism that overcomes limitations of existing robust aggregation algorithms.

Main Methods:

• Constructing an undirected weighted graph using a Gaussian kernel function to map data features into an infinite-dimensional Hilbert space.
• Employing normalized cut (NCut) on the graph's Laplacian matrix to partition and identify malicious clients.
• Developing a novel lens detection method to address similarities between benign and malicious updates.

Main Results:

• The proposed spectral clustering algorithm (SCA) demonstrates strong robustness against all tested attacks across IID and Non-IID settings.
• The method effectively identifies malicious updates that traditional robust aggregation algorithms fail to detect.
• Experimental results show superior performance compared to existing defense methods.

Conclusions:

• The enhanced spectral clustering algorithm provides an effective unsupervised defense against malicious clients in federated learning.
• The approach enhances model accuracy and convergence by accurately identifying and mitigating adversarial updates.
• This method offers a robust solution for securing federated learning systems against evolving threats.