ETD

• BFV
• BGV
• convolutional neural networks
• direct feedback alignment
• encrypted training
• exact ReLU
• homomorphic encryption
• neural networks
• no-prop
• OpenFHE
• privacy preserving deep learning
• TFHE

This thesis addresses the increasing privacy concerns associated with Artificial Neural Networks (ANNs) in sensitive domains such as healthcare and finance. It explores Homomorphic Encryption (HE) as a solution to perform inference on encrypted data without decryption, preserving confidentiality on cloud servers. The work provides a comparative analysis of primary HE schemes: BFV/BGV for exact integer arithmetic, CKKS for approximate real-number operations, and TFHE for boolean circuits. It highlights the critical challenge of noise accumulation and the computational cost of Bootstrapping required for fully homomorphic operations.
Investigating state-of-the-art frameworks, the study analyzes TFHE-NN for encrypted training and Orion for inference via CKKS. The core contribution is a novel implementation of encrypted inference using the BFV scheme within the OpenFHE library. This approach targets the creation of an exact ReLU activation function to overcome the approximation errors typical of Chebyshev polynomials used in CKKS. Two implementation strategies are proposed: a client-aided "online" protocol that delegates polynomial decomposition to the user to bypass bootstrapping, and an autonomous "leveled" approach that ensures full server-side security. Experimental results demonstrate the feasibility of exact integer-based inference for privacy-preserving neural networks.