Efficient CNN building blocks for encrypted data
Efficient CNN Building Blocks for Encrypted Data
AI summary
The paper proposes efficient CNN building blocks for encrypted data, using fully homomorphic encryption (FHE) to protect both input data and model parameters. Based on the CKKS scheme of the open-source HElib library, parameters such as the cyclotomic polynomial degree, depth limit, and computational precision significantly affect the choice of activation functions and pooling, creating trade-offs among precision, security level, and computation time. MNIST encrypted inference experiments show that ciphertext packing strategies and multithreaded parallelization are key to determining throughput and latency.