Efficient Keyset Design for Neural Networks Using Homomorphic Encryption

Area of Science:

• Cryptography
• Machine Learning
• Data Security

Background:

• Internet of Things (IoT) generates sensitive data, increasing reliance on Machine Learning as a Service (MLaaS).
• Growing privacy concerns necessitate Privacy-Preserving Machine Learning (PPML).
• Fully Homomorphic Encryption (FHE) enables computation on encrypted data but faces efficiency challenges.

Purpose of the Study:

• To address the computational overhead in FHE-based neural network inference.
• To optimize the rotation keyset design for improved FHE efficiency.
• To reduce memory consumption and computational costs in PPML.

Main Methods:

• Systematic exploration of three key design spaces (KDS) for rotation keyset optimization.
• Development of an optimized rotation keyset.
• Evaluation through two case studies demonstrating memory and speed improvements.

Main Results:

• Achieved up to 11.29x memory reduction.
• Demonstrated 1.67x to 2.55x speedup in FHE-based neural network inference.
• Validated the effectiveness of the proposed KDS design.

Conclusions:

• Optimizing rotation keyset design is a viable strategy to enhance FHE efficiency for PPML.
• The proposed KDS approach offers significant memory and computational benefits.
• This work contributes to more practical and efficient privacy-preserving machine learning solutions.