Swarm-FHE: Fully Homomorphic Encryption-based Swarm Learning for Malicious Clients

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Distributed Systems

Background:

• Swarm Learning (SL) enables collaborative model training without a central server, but faces privacy challenges due to data sensitivity and potential gradient leakage from neural networks like GANs.
• Existing SL frameworks use blockchain for secure aggregation, yet remain vulnerable to compromised participants who can manipulate others' privacy.
• The risk of original data reconstruction from model parameters necessitates advanced privacy-preserving techniques in distributed learning.

Purpose of the Study:

• To propose and evaluate Swarm-FHE, a novel method integrating Swarm Learning with Fully Homomorphic Encryption (FHE) to address privacy concerns in collaborative AI training.
• To enhance the security of distributed model training against malicious participants and gradient leakage.
• To leverage blockchain for participant authentication and secure parameter sharing in an encrypted environment.

Main Methods:

• Implemented Swarm-FHE, a framework that encrypts model parameters using Fully Homomorphic Encryption (FHE) before sharing them among participants in a Swarm Learning environment.
• Utilized blockchain technology for secure registration and authentication of participants within the Swarm Learning network.
• Evaluated the method's effectiveness through training convolutional neural networks on CIFAR-10 and MNIST datasets, testing various hyperparameter settings.

Main Results:

• Swarm-FHE successfully encrypts model parameters, safeguarding against gradient leakage and privacy manipulation by malicious participants.
• The proposed method demonstrated superior performance compared to existing approaches in secure collaborative training scenarios.
• Experiments on CIFAR-10 and MNIST datasets validated the efficacy and robustness of Swarm-FHE under different configurations.

Conclusions:

• Swarm-FHE offers a robust solution for privacy-preserving collaborative AI training by combining Swarm Learning, Fully Homomorphic Encryption, and blockchain technology.
• The integration of FHE effectively mitigates risks associated with data sensitivity and malicious actors in distributed learning environments.
• This approach significantly advances the security and privacy standards for decentralized machine learning applications.