Robust Aggregation for Federated Learning by Minimum γ-Divergence Estimation.

Area of Science:

• Machine Learning
• Distributed Systems
• Data Security

Background:

• Federated learning enables collaborative model training without data sharing.
• Traditional aggregation methods like sample mean are vulnerable to outliers and Byzantine attacks.
• Existing robust methods include median and trimmed-mean, but a novel approach is needed.

Purpose of the Study:

• To introduce a novel robust aggregation method, gamma-mean, for federated learning.
• To enhance the resilience of federated learning against malicious clients (Byzantine problem).
• To provide a data-driven weighting scheme for improved aggregation.

Main Methods:

• Proposing gamma-mean, a minimum divergence estimation based on density power divergence.
• Implementing a data-driven weighting scheme controlled by the gamma value.
• Analyzing robustness using influence function and presenting numerical results.

Main Results:

• Gamma-mean effectively mitigates the influence of Byzantine clients.
• The proposed method demonstrates robustness against malicious data poisoning.
• Numerical results validate the effectiveness of gamma-mean.

Conclusions:

• Gamma-mean offers a robust and effective aggregation strategy for federated learning.
• This method enhances security and reliability in collaborative model training.
• The data-driven weighting scheme provides tunable robustness.