Privacy-Preserving Neural Network Inference Framework via Homomorphic Encryption and SGX

Xiao, Huizi; Zhang, Qingyang; Pei, Qingqi; Shi, Wei

doi:10.1109/icdcs51616.2021.00077

Cited by 6 publications

(1 citation statement)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To circumvent the need for a TA, an alternative approach is to employ a trusted execution environment (TEE) like Intel SGX [37] and AMD SEV [38]. For instance, Xiao et al [39] and Takeshita et al [40] employed the TEE of Intel SGX to manage FHE key pairs, combining plaintext execution in the TEE with ciphertext execution using FHE in the rich execution environment (REE). Yakupoglu et al [41] developed three secure multi-party computation (SMPC) protocols leveraging the TEE of Intel SGX and advanced encryption standards (AES).…”

Section: Discussionmentioning

confidence: 99%

Privacy Preserving Function Evaluation Using Lookup Tables with Word-Wise FHE

LI,

YAMANA

2024

IEICE Trans. Fundamentals

View full text Add to dashboard Cite

Homomorphic encryption (HE) is a promising approach for privacy-preserving applications, enabling a third party to assess functions on encrypted data. However, problems persist in implementing privacy-preserving applications through HE, including 1) long function evaluation latency and 2) limited HE primitives only allowing us to perform additions and multiplications. A homomorphic lookup-table (LUT) method has emerged to solve the above problems and enhance function evaluation efficiency. By leveraging homomorphic LUTs, intricate operations can be substituted. Previously proposed LUTs use bit-wise HE, such as TFHE, to evaluate single-input functions. However, the latency increases with the bit-length of the function's input(s) and output. Additionally, an efficient implementation of multi-input functions remains an open question. This paper proposes a novel LUT-based privacy-preserving function evaluation method to handle multi-input functions while reducing the latency by adopting word-wise HE. Our optimization strategy adjusts table sizes to minimize the latency while preserving function output accuracy, especially for common machine-learning functions. Through our experimental evaluation utilizing the BFV scheme of the Microsoft SEAL library, we confirmed the runtime of arbitrary functions whose LUTs consist of all input-output combinations represented by given input bits: 1) single-input 12-bit functions in 0.14 s, 2) single-input 18-bit functions in 2.53 s, 3) two-input 6-bit functions in 0.17 s, and 4) three-input 4-bit functions in 0.20 s, employing four threads. Besides, we confirmed that our proposed table size optimization strategy worked well, achieving 1.2 times speed up with the same absolute error of order 10 −4 for Swish and 1.9 times speed up for ReLU while decreasing the absolute error from order 10 −2 to 10 −4 compared to the baseline, i.e., polynomial approximation.

show abstract

Section: Discussionmentioning

confidence: 99%