Practical Privacy-Preserving Machine Learning using Fully Homomorphic Encryption¶
Year: 2023
type: technique collection: ppml status: auto
Summary¶
Auto-extracted summary (unreviewed)
Presents a practical approach to training a binary classifier (a linear SVM with added non-linear features) over CKKS-encrypted data using a Newton-based optimizer implemented in SEAL, with a three-actor model of Data Owners who encrypt samples, a Client, and a Cloud that jointly run training. On the WBC breast-cancer dataset the training phases complete in under 45 s (43.6 s excluding encryption, 100.3 s including it) on a mid-range computer at 128-bit security, each Data Owner spends a constant ~150 ms per sample, and the method scales to 8,000 samples and 60 features.
Abstract¶
Machine learning is a widely-used tool for analysing large datasets, but increasing public demand for privacy preservation and the corresponding introduction of privacy regulations have severely limited what data can be analysed, even when this analysis is for societal benefit. Homomorphic encryption, which allows computation on encrypted data, is a natural solution to this dilemma, allowing data to be analysed without sacrificing privacy. Because homomorphic encryption is computationally expensive, however, current solutions are mainly restricted to use it for inference and not training. In this work, we present a practically viable approach to privacypreserving machine learning training using fully homomorphic encryption. Our method achieves fast training speeds, taking less than 45 seconds to train a binary classifier over thousands of samples on a single mid-range computer, significantly outperforming state-of-the-art results.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | Fast FHE-based training of a binary SVM classifier | llm | 0.85 |
| contribution | Newton-based training algorithm evaluated under CKKS | llm | 0.85 |
| contribution | Column-oriented ciphertext packing exploiting CKKS SIMD slots | llm | 0.85 |
| contribution | Scalability analysis over number of samples and features | llm | 0.85 |
| schemes | CKKS | llm | 0.85 |
| libraries | SEAL | llm | 0.85 |
| libraries | SEAL-Python | llm | 0.85 |
| models | linear SVM (with non-linear features) | llm | 0.85 |
| datasets | WBC (breast cancer) | llm | 0.85 |
| tasks | binary classification | llm | 0.85 |
| tasks | privacy-preserving training | llm | 0.85 |
| hardware | single mid-range personal computer | llm | 0.85 |
| security level | 128-bit equivalent | llm | 0.85 |
| threat model | outsourced training among Data Owners, Client and Cloud | llm | 0.85 |
| interaction | interactive (HE+MPC) | llm | 0.85 |
| code url | https://github.com/Huelse/SEAL-Python (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/Microsoft/SEAL.Microsoft (URL found in the full text) | regex | 0.50 |
| limitations | A linear SVM, even with non-linear features, cannot handle complex datasets such as MNIST | llm | 0.85 |
| limitations | Beyond 8,192 samples per ciphertext, training time grows linearly (though parallelisable) | llm | 0.85 |
| limitations | Data ingest time grows linearly with the number of samples | llm | 0.85 |
| llm relevance | low | llm | 0.85 |
| key result | training time without encryption: 43.6 s (WBC dataset, 80/20 split, data ingest + initialisation + optimisation) | llm | 0.85 |
| key result | total time with encryption: 100.3 s (WBC dataset, encryption 56.7 s) | llm | 0.85 |
| key result | encryption time per Data Owner: about 150 ms per sample (constant) (up to 8,000 samples) | llm | 0.85 |
| key result | scalability: up to 8,000 samples and 60 features (training time constant up to 8,192 samples per ciphertext) | llm | 0.85 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 19 [p.4], MPC 12 [p.4], Partially homomorphic encryption 11 [p.4], RSA 3 [p.6], TFHE 2 [p.1], Garbled circuits 2 [p.3]
- library: Microsoft SEAL 8 [p.10], HElib 2 [p.4], Concrete (Zama) 2 [p.1], nGraph-HE 2 [p.4]
- model: SVM 18 [p.1], Logistic regression 1 [p.13]
- dataset: MNIST 4 [p.3], Breast cancer (Wisconsin) 4 [p.5], UCI repository 1 [p.14]
- hardware: GPU 2 [p.5]
- technique: Polynomial approximation 15 [p.4], Chebyshev 11 [p.4], Newton iteration 10 [p.5], ReLU 10 [p.4], SIMD packing 6 [p.5], Bootstrapping 5 [p.4], Softmax 4 [p.8], Sigmoid 4 [p.4], RNS 2 [p.4], Federated learning 2 [p.1], Encrypted training 2 [p.5], Compiler 2 [p.4]
- assumption: Lattice-based 3 [p.6], RLWE 2 [p.6], Post-quantum 2 [p.4]
- task: Medical / healthcare 28 [p.1], Finance 5 [p.1]
- topic: Homomorphic encryption 78 [p.1], Privacy-preserving machine learning 39 [p.1], Private / secure inference 3 [p.1], Open source 3 [p.10]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/inference/brand-2023_practical-privacy-preserving-ml-using-fhe.pdf |
ppml/inference | 15 | 836.3 KB | 3d6ca6faabee |
was Brand and Pradel - 2023 - Practical Privacy-Preserving Machine Learning using Fully Homomorphic Encryption.pdf |
Database key: practical-privacy-preserving-machine-learning-using-fully-homomorphic · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set practical-privacy-preserving-machine-learning-using-fully-homomorphic COLUMN VALUE