Paper 2026/1381

PriFT: Private Fine-Tuning using off-the-shelf MPC and HE libraries

Abstract

Privacy-Preserving Machine Learning (PPML) is a methodology designed to maintain data privacy throughout the machine learning pipeline. Although cryptographically protecting input privacy at the training phase is theoretically feasible, it remains computationally intensive in practice. As such, most recent works in this area focus on the inference phase. In contrast, we consider the training phase. Our goal is to enable machine learning engineers to use customer data earlier in the ML pipeline without compromising customer privacy or violating regulations. In this work, we introduce a framework named PriFT (Private Fine-Tuning), which leverages a transformer as a feature extractor and then performs training of a neural network on privacy-protected features. PriFT supports fully-private training, where the data is encrypted in the entire ML pipeline, as well as semi-private training, which balances privacy and performance by decrypting the true and predicted labels during training. PriFT can perform secure training both by means of Multiparty Computation (MPC) and based on Homomorphic Encryption (HE), which allows for a direct comparison of the two most prevalent cryptographic solutions for secure computation on a real-world use case. The codebase of our experiments is fully open-sourced and based on well-established libraries, namely Crypten and TenSEAL. Our experimental results show that the MPC approach largely outperforms the HE approach, especially in the semi-private setting. Furthermore, the MPC-based solution in the semi-private setting outperforms the fully-private training approximately by 3×, offering an adequate privacy-performance tradeoff. Our results show that both HE and MPC can achieve accuracy close to that of plaintext models.