An Adaptive Model Filtering Algorithm Based on Grubbs Test in Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) enables decentralized model training while preserving data privacy.
• FL systems are vulnerable to poisoning attacks, degrading model performance and reliability.
• Existing defenses struggle to balance robustness and efficiency, particularly with non-independent and identically distributed (non-IID) data.

Purpose of the Study:

• To propose FedGaf, an adaptive model filtering algorithm for federated learning.
• To enhance the robustness and training efficiency of FL against poisoning attacks.
• To address the trade-off challenges in defending FL models, especially on non-IID datasets.

Main Methods:

• Developed an adaptive model filtering algorithm based on the Grubbs test (FedGaf).
• Designed multiple child adaptive model filtering algorithms to optimize robustness and efficiency.
• Implemented a dynamic decision mechanism using global model accuracy to minimize computational overhead.
• Incorporated a global model weighted aggregation method to accelerate model convergence.

Main Results:

• FedGaf demonstrates superior robustness and efficiency compared to existing methods.
• The proposed algorithm effectively defends against various poisoning attack strategies.
• Experimental results validate FedGaf's performance on both IID and non-IID datasets.
• FedGaf achieves a better trade-off between robustness and efficiency than prior defenses.

Conclusions:

• FedGaf offers an effective solution for defending federated learning against poisoning attacks.
• The algorithm provides a significant improvement in robustness and efficiency, particularly for non-IID data scenarios.
• FedGaf represents a promising advancement in securing federated learning systems.