Federated learning with bilateral defense via blockchain.

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Distributed Systems

Background:

• Federated Learning (FL) enhances data privacy but faces security threats like data breaches and poisoning attacks.
• Existing FL solutions often address only one security issue, leaving a gap in comprehensive defense.
• A non-fully trusted model with malicious clients and honest-but-curious servers presents a complex security challenge.

Purpose of the Study:

• To propose a robust Federated Learning scheme, FedBASS, that addresses both privacy breaches and model poisoning attacks.
• To develop a system resilient to non-IID data and secure against both malicious clients and servers.
• To ensure transparency, robustness, and practicality in Federated Learning environments.

Main Methods:

• Implemented a dual-server architecture (Analyzer and Verifier) with CKKS encryption for gradient security.
• Utilized cosine similarity for detecting malicious clients and a gradient compensation strategy for non-IID data.
• Integrated a weakened differential privacy scheme with shuffling for privacy during clustering and employed blockchain for communication integrity.

Main Results:

• FedBASS effectively mitigates privacy breaches and poisoning attacks in a challenging threat model.
• The scheme demonstrates robustness against non-IID data through dynamic clustering and gradient compensation.
• Blockchain integration ensures transparency and prevents selfish behaviors, enhancing overall system integrity.

Conclusions:

• FedBASS successfully balances model fidelity, robustness, efficiency, and privacy in Federated Learning.
• The proposed scheme offers a practical and secure solution for complex, untrusted environments.
• This work advances Federated Learning security by addressing multiple threats simultaneously.