Robust Federated Learning: Maximum Correntropy Aggregation Against Byzantine Attacks.

Area of Science:

• Decentralized Machine Learning
• Cybersecurity in AI

Background:

• Federated learning enables collaborative training while preserving participant privacy and security.
• Untrustworthy devices, such as Byzantine attackers, can compromise federated learning by uploading malicious parameters, corrupting the global model.

Purpose of the Study:

• To propose a novel robust aggregation method, Maximum Correntropy Aggregation (MCA), to defend federated learning against malicious parameter uploads.
• To utilize the Maximum Correntropy Criterion (MCC) as a similarity metric for robust parameter aggregation, unlike its previous use in denoising.

Main Methods:

• Developed MCA, applying MCC to derive a central value from parameters by measuring parameter distribution similarity.
• Employed fixed-point iteration to solve the optimization objective, demonstrating linear convergence.
• Conducted theoretical analysis to establish MCA's robustness aggregation property and error bounds.

Main Results:

• MCA effectively measures parameter distribution using correntropy, capturing high-order statistical properties to resist attackers.
• The method does not require knowledge of the proportion of malicious attackers.
• Experimental results on IID and non-IID data across three datasets demonstrate MCA's significant robustness against mainstream attacks.

Conclusions:

• MCA provides a robust aggregation method for federated learning, effectively mitigating the impact of malicious participants.
• The proposed technique offers a strong defense against Byzantine attacks, outperforming existing methods in resilience.