Enhanced Security and Privacy via Fragmented Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data.
• FL faces challenges in balancing model accuracy with participant privacy and data security.
• Malicious participants can poison updates, compromising model integrity, while privacy measures can reduce accuracy.

Purpose of the Study:

• To propose a novel federated learning framework, Fragmented Federated Learning (FFL), that addresses the accuracy-privacy-security conflict.
• To develop mechanisms for enhanced privacy and robust security within the FL process.
• To maintain global model accuracy while mitigating risks from malicious actors and privacy breaches.

Main Methods:

• Participants randomly exchange and mix encrypted fragments of their model updates before aggregation.
• A lightweight protocol facilitates private exchange and mixing of encrypted update fragments.
• A reputation-based defense system is implemented to assess participant trustworthiness and update quality.

Main Results:

• FFL prevents semi-honest servers from executing privacy attacks by obscuring individual update origins.
• The framework effectively counters poisoning attacks through a reputation-based security mechanism.
• Experiments demonstrate that FFL reconstructs the global model without accuracy loss.

Conclusions:

• Fragmented Federated Learning (FFL) offers a viable solution to the accuracy-privacy-security trade-off in federated learning.
• The proposed privacy protocol and security defense are effective against common FL threats.
• FFL enables secure, private, and accurate collaborative model training.