Selective Trimmed Average: A Resilient Federated Learning Algorithm With Deterministic Guarantees on the Optimality...

Area of Science:

• Machine Learning
• Distributed Systems
• Cybersecurity

Background:

• Federated learning (FL) trains models across decentralized data sources without sharing raw data.
• Traditional FL relies on server-worker architectures, vulnerable to adversarial agents causing data or model poisoning.
• Existing resilient methods often assume a trusted central server, limiting their applicability.

Purpose of the Study:

• To introduce a novel resilient algorithm, selective trimmed average (SETA), for federated learning.
• To enhance the robustness of federated learning against malicious participants and corrupted data.
• To develop a method effective in both server-worker and shared memory architectures.

Main Methods:

• Proposing SETA, an algorithm that filters and combines exchanged parameters from agents.
• Mathematically proving SETA's resilience against data and local model poisoning attacks.
• Evaluating SETA's performance on the MNIST and multiclass weather dataset (MWD).

Main Results:

• SETA effectively mitigates the impact of misbehaving agents in federated learning.
• The algorithm demonstrates resilience against both data and local model poisoning.
• Numerical results validate the theoretical findings on benchmark datasets.

Conclusions:

• SETA offers a robust solution for secure federated learning, particularly in untrusted environments.
• The algorithm's applicability extends to shared memory architectures, removing the need for a trusted server.
• SETA represents a significant advancement in defending federated learning systems against adversarial threats.