Training of Classification Models via Federated Learning and Homomorphic Encryption

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Increasing use of social networks and data protection laws necessitate secure methods for training machine learning models.
• User data on devices can contain sensitive information, posing risks if leaked.
• Current methods for decentralized data training require enhanced privacy measures.

Purpose of the Study:

• To propose a novel protocol for training Multi-Layer Perceptron (MLP) neural networks.
• To combine federated learning with homomorphic encryption for robust data privacy.
• To ensure data remains secure and distributed across multiple clients during the training process.

Main Methods:

• Developed a protocol integrating federated learning and homomorphic encryption for MLP training.
• Conducted simulations using a multi-class classification dataset.
• Varied MLP architectures and the number of participating clients to test the protocol's efficacy.

Main Results:

• Validated the proposed protocol through extensive simulations.
• Presented performance metrics in both local and federated settings.
• Conducted a comparative analysis against existing methods.

Conclusions:

• The proposed protocol effectively preserves data privacy during federated MLP training.
• Formal analysis confirms the privacy guarantees under defined assumptions.
• The protocol offers significant added value compared to previous approaches in secure machine learning.