Comparison of Fully Homomorphic Encryption and Garbled Circuit Techniques in Privacy-Preserving Machine Learning Inference¶
Authors: Kalyan Cheerla; Lotfi Ben Othmane; Kirill Morozov
Year: 2025
Venue: 2025 IEEE Secure Development Conference (SecDev)
type: benchmark collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
This SecDev paper compares FHE and garbled circuits for two-party secure inference of a fixed two-layer neural network, implementing it with CKKS in Microsoft SEAL and with Intel Labs' TinyGarble2.0, and measuring output error, round-trip time, peak memory, communication volume and rounds under a semi-honest model. The authors find that the modular GC implementation runs faster with lower memory, while FHE offers non-interactive inference; the excerpt reports no concrete numbers.
Abstract¶
Machine Learning (ML) is making its way into fields such as healthcare, finance, and Natural Language Processing (NLP), and concerns over data privacy and model confidentiality continue to grow. Privacy-preserving Machine Learning (PPML) addresses this challenge by enabling inference on private data without revealing sensitive inputs or proprietary models. Leveraging Secure Computation techniques from Cryptography, two widely studied approaches in this domain are Fully Homomorphic Encryption (FHE) and Garbled Circuits (GC). This work presents a comparative evaluation of FHE and GC for secure neural network inference. A two-layer neural network (NN) was implemented using the CKKS scheme from the Microsoft SEAL library (FHE) and the TinyGarble2.0 framework (GC) by IntelLabs. Both implementations are evaluated under the semihonest threat model, measuring inference output error, roundtrip time, peak memory usage, communication overhead, and communication rounds. Results reveal a trade-off: modular GC offers faster execution and lower memory consumption, while FHE supports non-interactive inference.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | side-by-side FHE (CKKS/SEAL) and GC (TinyGarble2.0) implementations of a two-layer NN inference | llm | 0.80 |
| contribution | evaluation of error, round-trip time, memory, communication overhead and rounds | llm | 0.80 |
| contribution | open-source implementations and instrumented EMP-tool fork | llm | 0.80 |
| schemes | CKKS | llm | 0.80 |
| schemes | garbled circuits | llm | 0.80 |
| schemes | oblivious transfer | llm | 0.80 |
| libraries | SEAL | llm | 0.80 |
| libraries | TinyGarble2.0 | llm | 0.80 |
| libraries | EMP-tool | llm | 0.80 |
| libraries | EMP-OT | llm | 0.80 |
| models | two-layer neural network | llm | 0.80 |
| tasks | secure neural network inference | llm | 0.80 |
| security level | 128-bit | llm | 0.80 |
| threat model | semi-honest two-party (client holds input, server holds model) | llm | 0.80 |
| interaction | other | llm | 0.80 |
| code url | https://github.com/Microsoft/SEAL (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/IntelLabs/TinyGarble2.0 (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/IntelLabs/emp-ot (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/kalyancheerla/snni-fhe-gc | llm | 0.80 |
| limitations | ReLU is approximated by x^2 and sigmoid by a quadratic, which diverge significantly for large inputs | llm | 0.80 |
| limitations | evaluation limited to a fixed-parameter two-layer neural network | llm | 0.80 |
| llm relevance | low | llm | 0.80 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 17 [p.1], MPC 8 [p.1], Garbled circuits 6 [p.1], Oblivious transfer 4 [p.1], Secret sharing 2 [p.2]
- library: Microsoft SEAL 10 [p.1], EVA / CHET 1 [p.2], ABY / ABY3 1 [p.2], CrypTen 1 [p.2]
- model: CNN 5 [p.2], BERT 1 [p.2], GPT 1 [p.2], Logistic regression 1 [p.7]
- dataset: MNIST 1 [p.6], CIFAR-10 1 [p.6]
- technique: ReLU 11 [p.3], Sigmoid 11 [p.3], Rescaling 5 [p.3], Polynomial approximation 4 [p.2], Non-interactive 4 [p.1], Convolution 3 [p.6], SIMD packing 2 [p.4], Rotations 2 [p.3], Matrix multiplication 2 [p.4], Lookup table 2 [p.6], Two-party computation 2 [p.1], Semi-honest model 2 [p.2]
- assumption: 128-bit security 2 [p.3]
- task: Medical / healthcare 5 [p.1], Language modeling / generation 4 [p.1]
- topic: Homomorphic encryption 59 [p.1], Private / secure inference 14 [p.2], Privacy-preserving machine learning 11 [p.1], Benchmark 6 [p.2], Open source 6 [p.7]
Related papers¶
- COMPARISON OF FULLY HOMOMORPHIC ENCRYPTION AND GARBLED CIRCUITS APPROACHES IN PRIVACY-PRESERVING MACHINE LEARNING Kalyan Cheerla Thesis Prepared for the Degree of MASTER OF SCIENCE UNIVERSITY OF NORTH TEXAS July 2025 version of this paper — MSc thesis (UNT, 2025) behind the SecDev 2025 paper
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/surveys/cheerla-2025_comparison-of-fhe-and-garbled-circuit-techniques-in-ppml-inference.pdf |
ppml/surveys | 8 | 590.4 KB | d39141859ff7 |
was Cheerla et al. - 2025 - Comparison of Fully Homomorphic Encryption and Garbled Circuit Techniques in Privacy-Preserving Mach.pdf |
Database key: comparison-of-fully-homomorphic-encryption-and-garbled-circuit · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set comparison-of-fully-homomorphic-encryption-and-garbled-circuit COLUMN VALUE