Based on model randomization and adaptive defense for federated learning schemes.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated Learning (FL) enables collaborative model training without data sharing, but faces security vulnerabilities from client-side attacks.
• Attacks targeting model updates and malicious client behavior compromise FL integrity and privacy.

Purpose of the Study:

• To propose a novel Privacy-Preserving and High-Secure Federated Learning (PPHSFL) scheme.
• To enhance the security and privacy of FL systems against various threats.

Main Methods:

• Implemented Models Randomization and Compensation (MRC) using a randomized discrete loss function to prevent gradient backward inference.
• Introduced Adaptive Defensive Rewards (ADR) for adaptive client selection and dynamic rewards to counter dishonest clients.

Main Results:

• The PPHSFL scheme effectively preserves customer privacy during FL processing.
• Demonstrated significant mitigation of threats posed by malicious clients and model update attacks.
• Achieved an average accuracy improvement of 3.0% compared to state-of-the-art methods.

Conclusions:

• The PPHSFL scheme offers a robust solution for secure and private Federated Learning.
• MRC and ADR effectively address key vulnerabilities in FL, improving overall system security and performance.