FedID: Enhancing Federated Learning Security Through Dynamic Identification.

Area of Science:

• Machine Learning
• Cybersecurity
• Decentralized Systems

Background:

• Federated learning (FL) offers privacy but is vulnerable to backdoor attacks.
• Current defenses struggle against stealthy attacks with non-IID data or similar gradients.

Purpose of the Study:

• To develop a robust defense against diverse backdoor attacks in FL.
• To address limitations of existing distance-based defenses in high dimensions and non-IID settings.

Main Methods:

• Propose FedID, a strategy using multiple metrics with dynamic weighting for backdoor detection.
• Introduce a modified z-score for selecting gradients during aggregation.
• Evaluate FedID against various attacks and non-IID data distributions.

Main Results:

• FedID effectively detects diverse malicious gradients, outperforming prior defenses.
• Demonstrates robustness against adaptive attacks designed to bypass defenses.
• Maintains high performance on benign tasks across various non-IID data distributions.

Conclusions:

• FedID offers a simple, effective, and adaptive solution for backdoor defense in FL.
• The multi-metric approach overcomes limitations of single-metric and Euclidean distance-based methods.
• FedID is resilient to adaptive attacks and performs well under non-IID conditions.