Paper 2024/1859

Non-interactive Fully Encrypted Machine Learning Protocol for Inference

Abstract

As privacy concerns have arisen in machine learning, privacy-preserving machine learning (PPML) has received significant attention. Fully homomorphic encryption (FHE) and secure multi-party computation (MPC) are representative building blocks for PPML. However, in PPML protocols based on FHE and MPC, interaction between the client (who provides encrypted input data) and the evaluator (who performs the computation) is essential to obtain the final result in plaintext. Functional encryption (FE) is a promising candidate to remove this constraint, but existing FE-based PPML protocols are restricted to evaluating only simple ML models, such as one-layer neural networks, or they support partially encrypted PPML, which makes them vulnerable to information leakage beyond the inference results. In this paper, we propose a fully encrypted FE-based PPML protocol, which supports the evaluation of arbitrary functions over encrypted data with no information leakage during computation, for the first time. To achieve this, {we introduce a new functional encryption scheme for quadratic polynomials, in which the encryption algorithm is realized as a linear function with respect to a message. This structural property allows for the composition of the encryption algorithm with arbitrary quadratic functions, thereby enabling multiple compositions of quadratic polynomials to compute arbitrary complex functions in an encrypted manner. Our FE-based PPML protocol is secure in the malicious model, which means that an adversary cannot obtain any information about the input data even though they intentionally deviate from the protocol. We then show how to use our protocol to build a fully encrypted 2-layer neural network model with quadratic activation functions and present experimental results.

Note: change the title, and description