Fully Homomorphically Encrypted Deep Learning as a Service¶
Authors: George Onoufriou; Paul Mayfield; Georgios Leontidis
Year: 2021
Venue: Machine Learning and Knowledge Extraction
type: application collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
The authors build a two-part client/server 'Encrypted Deep Learning as a Service' (EDLaaS) pipeline on Microsoft SEAL via their own Python bindings (ReSeal) and apply it to encrypted inference with a 1D CNN for milk-yield sequence prediction in the agri-food sector. Training is done in plaintext; encrypted inference reproduces the plaintext loss to three significant figures, giving a MAPE of 12.4% (87.6% accuracy), while ciphertexts are far larger than plaintexts (9.60 MB vs 0.0656 MB at polynomial modulus degree 16384). They discuss privacy limitations of plaintext training and data-sharing implications for the agri-food supply chain.
Abstract¶
Fully Homomorphic Encryption (FHE) is a relatively recent advancement in the field of privacy-preserving technologies. FHE allows for the arbitrary depth computation of both addition and multiplication, and thus the application of abelian/polynomial equations, like those found in deep learning algorithms. This project investigates how FHE with deep learning can be used at scale toward accurate sequence prediction, with a relatively low time complexity, the problems that such a system incurs, and mitigations/solutions for such problems. In addition, we discuss how this could have an impact on the future of data privacy and how it can enable data sharing across various actors in the agri-food supply chain, hence allowing the development of machine learning-based systems. Finally, we find that although FHE incurs a high spatial complexity cost, the run time is within expected reasonable bounds, while allowing for absolutely private predictions to be made, in our case for milk yield prediction with a Mean Absolute Percentage Error (MAPE) of 12.4% and an accuracy of 87.6% on average.
Source: crossref
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | client-server EDLaaS system for encrypted inference at scale (containerised) | llm | 0.80 |
| contribution | FHE-compatible 1D CNN for sequence (milk yield) prediction | llm | 0.80 |
| contribution | measurement of encrypted vs plaintext time, accuracy and ciphertext size | llm | 0.80 |
| contribution | open-source Python bindings to MS-SEAL (python-reseal) | llm | 0.80 |
| contribution | discussion of privacy and data sharing in the agri-food supply chain | llm | 0.80 |
| schemes | CKKS | llm | 0.80 |
| libraries | SEAL | llm | 0.80 |
| libraries | python-reseal (ReSeal) | llm | 0.80 |
| libraries | NumPy | llm | 0.80 |
| models | 1D CNN | llm | 0.80 |
| datasets | milk-yield time series (17,309 test examples) | llm | 0.80 |
| tasks | milk yield forecasting (sequence regression) | llm | 0.80 |
| tasks | encrypted deep learning inference as a service | llm | 0.80 |
| threat model | client-server; plaintext training, encrypted inference | llm | 0.80 |
| interaction | non-interactive FHE | llm | 0.80 |
| code url | https://github.com/DreamingRaven/python-reseal (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/DreamingRaven/python-reseal | llm | 0.80 |
| limitations | high spatial (ciphertext size) cost, partly due to serialisation between C++ SEAL and Python | llm | 0.80 |
| limitations | training/backpropagation is done in plaintext, so model weights can approximate the data distribution | llm | 0.80 |
| limitations | inefficient custom MS-SEAL bindings limit speed | llm | 0.80 |
| limitations | some low-level remote operations left unimplemented | llm | 0.80 |
| limitations | evaluation only on a LAN at small scale relative to production | llm | 0.80 |
| llm relevance | medium | llm | 0.80 |
| key result | MAPE: 12.4% (accuracy 87.6%) (encrypted 1D CNN milk yield prediction, test set of 17,309 examples) | llm | 0.80 |
| key result | encrypted vs plaintext loss difference: negligible (identical to 3 significant figures) (same network, plaintext vs ciphertext inference) | llm | 0.80 |
| key result | ciphertext size: 9.60 MB vs 0.0656 MB plaintext (polynomial modulus degree 16384, 8192 slots) | llm | 0.80 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 8 [p.2], BFV 3 [p.2]
- library: Microsoft SEAL 9 [p.2], PALISADE 1 [p.2], EVA / CHET 1 [p.8]
- model: CNN 24 [p.1], Logistic regression 1 [p.16], Autoencoder 1 [p.15]
- dataset: MNIST 2 [p.2], CIFAR-10 2 [p.2], CIFAR-100 1 [p.2]
- hardware: NVIDIA Jetson 2 [p.7], GPU 1 [p.8]
- technique: Sigmoid 20 [p.1], Convolution 9 [p.1], ReLU 8 [p.1], Bootstrapping 7 [p.8], Softmax 3 [p.2], Modulus switching 2 [p.12]
- assumption: RLWE 5 [p.1], LWE 3 [p.1]
- task: IoT / edge 3 [p.3]
- topic: Homomorphic encryption 64 [p.1], Open source 4 [p.6]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/inference/onoufriou-2021_fully-homomorphically-encrypted-deep-learning-as-a-service.pdf |
ppml/inference | 16 | 879.4 KB | f8737ed6f4a0 |
was Onoufriou et al. - 2021 - Fully Homomorphically Encrypted Deep Learning as a Service.pdf |
Database key: fully-homomorphically-encrypted-deep-learning-as-a-service · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set fully-homomorphically-encrypted-deep-learning-as-a-service COLUMN VALUE