Privacy-preserving neural networks with Homomorphic encryption: Challenges and opportunities

Abstract

Classical machine learning modeling demands considerable computing power for internal calculations and training with big data in a reasonable amount of time. In recent years, clouds provide services to facilitate this process, but it introduces new security threats of data breaches. Modern encryption techniques ensure security and are considered as the best option to protect stored data and data in transit from an unauthorized third-party. However, a decryption process is necessary when the data must be processed or analyzed, falling into the initial problem of data vulnerability. Fully Homomorphic Encryption (FHE) is considered the holy grail of cryptography. It allows a non-trustworthy third-party resource to process encrypted information without disclosing confidential data. In this paper, we analyze the fundamental concepts of FHE, practical implementations, state-of-the-art approaches, limitations, advantages, disadvantages, potential applications, and development tools focusing on neural networks. In recent years, FHE development demonstrates remarkable progress. However, current literature in the homomorphic neural networks is almost exclusively addressed by practitioners looking for suitable implementations. It still lacks comprehensive and more thorough reviews. We focus on the privacy-preserving homomorphic encryption cryptosystems targeted at neural networks identifying current solutions, open issues, challenges, opportunities, and potential research directions.

Privacy–Enhancing Face Biometrics: A Comprehensive Survey

“…Homomorphic encryption schemes can in general be partitioned into three main sub-groups, i.e. [248], [257]:…”

“…Homomorphic encryption is relevant also beyond the field of biometrics. The reader is referred to some of the existing surveys, e.g., [257], [265], [266] for more detailed information on this topic.…”

Toward Learning Trustworthily from Data Combining Privacy, Fairness, and Explainability: An Application to Face Recognition

“…Consequently HE, contrarily to anonymization and perturbation techniques, entirely preserves utility and, contrarily to other simple distributed protocols techniques, automatically guarantees preserving the privacy of the single individuals. The major drawback of HE is its high computational overhead and the limitations for some operations [72,73]. In particular, three possible approaches are defined: Partially, Somewhat, and Fully HE.…”

“…For our purpose, namely ML-related applications, Somewhat HE is the most exploited approach, since it delivers the best trade-off [71] for our application. HE, on one hand, allows ensuring the privacy of the single individual, especially in the commonly adopted case where the computing and memory resources are outsourced to a third-party service provider, but on the other hand dramatically increases the computational requirements and reduces the possible network architectural choices [72,73]. Algorithmic Fairness deals with the problem of ensuring that the learned ML does not discriminate subgroups in the population using pre-in-and post-processing methods [26,74].…”

Framework for Privacy-Preserving Wearable Health Data Analysis: Proof-of-Concept Study for Atrial Fibrillation Detection

“…Numerous open-source homomorphic encryption libraries have arisen in recent years to facilitate the development and use of homomorphic encryption for AI models, including SEAL, HElib, TFHE, PALISADE, cuHE, HEAAN, and HE-transformer [16]. Microsoft's Simple Encrypted Arithmetic Library (SEAL) [17], with support for the Brakerski/Fan-Vercauteren (BFV) [18] and Cheon-Kim-Kim-Song (CKKS) [19] schemes, and IBM's HE-Lib [20] based on the Brakerski-Gentry-Vaikuntanathan (BGV) [21] scheme, are two of the most widely used libraries for building privacy-preserving neural network inference applications.…”