PREMAL — Privacy-Preserving ML with FHE
A curated map of papers on machine learning under Fully Homomorphic Encryption. Each entry is summarized into a per-paper page with pipeline and architecture diagrams, threat model, results, and limitations.
Scope: end-to-end FHE for ML — inference, training, federated aggregation. Out of scope: pure MPC, pure differential privacy, and trusted-execution-environment work. Hybrid schemes are in.
At a glance
Featured
Practical and Private Hybrid ML Inference with Fully Homomorphic Encryption 2025
Recent advances of privacy-preserving machine learning based on (Fully) Homomorphic Encryption 2024
Privacy-Preserving Machine Learning With Fully Homomorphic Encryption for Deep Neural Network 2022
A Survey of Deep Learning Architectures for Privacy-Preserving Machine Learning With Fully Homomorphic Encryption 2022
Towards the AlexNet Moment for Homomorphic Encryption: HCNN, the First Homomorphic CNN on Encrypted Data With GPUs 2021
A Survey on Privacy-Preserving Machine Learning with Fully Homomorphic Encryption 2020
All papers
Sorted by year (newest first), then by single-inference latency (ascending). Hover the inference time to see the hardware it was measured on.
| Title | Year | Hook | FHE | Model | Dataset | Accuracy | Inference (s) | Code |
|---|---|---|---|---|---|---|---|---|
| Touch of Privacy: A Homomorphic Encryption-Powered Deep Learning Framework for Fingerprint Authentication FHE fingerprint matching |
2025 | CKKS-encrypted Euclidean matching of CNN fingerprint features hits 99.06% accuracy in 0.025 s per comparison. | CKKS | CNN | SOCOFing | 99.06% (EER 0.40%, TAR 99.19%, FAR 0%) | 0.025 | — |
| Evaluation of Privacy-aware Support Vector Machine (SVM) Learning using Homomorphic Encryption CKKS SVM parameter sweep |
2025 | Parameter sweep of CKKS-encrypted SVM-Linear vs SVM-Poly on Iris: ring dimension and modulus dominate cost. | CKKS | SVM | Iris | 96.7% (matches plaintext) | 0.2029 | — |
| Evaluation of Privacy-Preserving Support Vector Machine (SVM) Learning Using Homomorphic Encryption CKKS SVM parameter sweep |
2025 | Parameter sweep of CKKS-encrypted SVM inference on Iris using OpenFHE; ring dimension and modulus dominate cost. | CKKS | SVM | Iris | 96.7% (0.0 vs plaintext) | 0.2029 | GitHub |
| Application of Machine Learning-NLP Approach With Fully Homomorphic Encryption Techniques in Medical PII Data FHE on medical PII |
2025 | Benchmarks Concrete-ML SGD, XGBoost, and Decision Tree on synthetic medical PII text classification. | TFHE (via Concrete-ML) | SGD / XGBoost / Decision Tree | Synthetic GAN-generated medical PII (100k) | 100% (SGD, XGB); 91.7% (DT) — matches plaintext | 0.41 | — |
| Comparison of Fully Homomorphic Encryption and Garbled Circuits Approaches in Privacy-Preserving Machine Learning FHE vs GC apples-to-apples |
2025 | Side-by-side benchmark: CKKS FHE vs TinyGarble2.0 GC on the same 2-layer NN. | CKKS (vs Garbled Circuits) | FCNN | Synthetic 3-dim input vectors | Max deviation: GC 23.46%, FHE 121.69% vs plaintext | 5.07723 | GitHub |
| Comparison of Fully Homomorphic Encryption and Garbled Circuit Techniques in Privacy-Preserving Machine Learning Inference FHE vs GC head-to-head |
2025 | Head-to-head SEAL-CKKS vs TinyGarble2.0 on a 2-layer NN — GC ~161x slower, FHE ~20,912x slower than plaintext. | CKKS | FCNN | Synthetic 3-dim inputs | Output deviation up to 121.69% (FHE) vs 23.46% (GC) from plaintext | 5.07723 | GitHub |
| Practical and Private Hybrid ML Inference with Fully Homomorphic Encryption Hybrid client/server FHE |
2025 | Hybrid FHE inference: server runs encrypted linear layers, client runs plaintext ReLU, no bootstrapping. | TFHE-based RLWE (hybrid client/server) | ResNet-20 / ResNet-18 / ResNet-34 | Multiple | 90.00% on CIFAR-10 (-1.70 vs ORION) | 13.65 | — |
| A privacy preserving machine learning framework for medical image analysis using quantized fully connected neural networks with TFHE based inference 4-bit QAT FCNN on MedMNIST |
2025 | Quantization-aware FCNN runs encrypted MedMNIST inference under TFHE with near-plaintext accuracy. | TFHE | FCNN | MedMNIST (Pneumonia/Breast/Blood) | 88.88% Pneumonia (-0.01 vs plaintext) | 119.55 | — |
| Activate Me!: Designing Efficient Activation Functions for Privacy-Preserving Machine Learning with Fully Homomorphic Encryption Scheme-switching ReLU |
2025 | Compares square, Chebyshev-50 poly-ReLU, and a CKKS<->TFHE scheme-switching ReLU on LeNet-5 and ResNet-20. | hybrid (CKKS+TFHE) | LeNet-5 and ResNet-20 | MNIST + CIFAR-10 | ResNet-20 89.8% (-2.4 vs plaintext) with scheme-switching; 83.8% (-8.4) with poly-approx | 1145.0 | — |
| An Efficient and Privacy-Preserving Federated Learning Approach Based on Homomorphic Encryption DREL + BFV FedAvg |
2025 | Dynamic per-layer quantization enables BFV-encrypted FedAvg with 192-bit security and >99% encryption-time savings vs Paillier. | BFV | CNN + FNN | Multiple | Up to +0.24% vs No-PPFL; worst -0.39% vs CKKS (FMNIST) | — | — |
| HHEML: Hybrid Homomorphic Encryption for Privacy-Preserving Machine Learning on Edge FPGA Pasta HHE |
2025 | First end-to-end FPGA HHE pipeline: dual-XOF Pasta accelerator cuts MNIST client-side latency 50x. | hybrid (Pasta SE + FHE via GuardML) | N/A | MNIST | Not reported | — | — |
| Federated Learning: An approach with Hybrid Homomorphic Encryption Hybrid HE Federated |
2025 | First end-to-end HHE-FL framework pairing PASTA cipher with BFV, cutting client uploads 2000x. | hybrid (PASTA + BFV) | CNN | MNIST | 97.6% (-1.3 vs plaintext 98.9%) | — | — |
| Swift: Fast Secure Neural Network Inference With Fully Homomorphic Encryption FHE-SS exact ReLU |
2025 | FHE+SS hybrid hides ReLU/max-pool, beating Cheetah 7.4x on ReLU under WAN. | CKKS | CNN | MNIST | Not reported | — | — |
| BlindFL: Segmented Federated Learning with Fully Homomorphic Encryption Segmented FHE FedAvg |
2025 | Segmented FedAvg halves CKKS aggregation time and blunts client-side gradient-inversion attacks. | CKKS | LeNet-5 / ResNet-20 | MNIST + CIFAR-10 | MNIST 98.53% / CIFAR-10 78.69% (vs 98.71% / 79.92% standard FL) | — | — |
| The Beginner’s Textbook for Fully Homomorphic Encryption Beginner FHE textbook |
2025 | Self-contained mathematical textbook deriving TFHE, BFV, CKKS, BGV and their RNS variants from first principles. | TFHE, BFV, CKKS, BGV | N/A | N/A | N/A | — | GitHub |
| Building Privacy-Preserving AI system: AI Inference on Encrypted Data with Fully Homomorphic Encryption Concrete ML production study |
2025 | Industrial feasibility study of Zama Concrete ML for TFHE inference on real-world XGBoost binary classifiers. | TFHE | XGBoost / Random Forest (binary classifier) | Proprietary binary-classification (3000 records) | 99% agreement (Model A); 22% agreement (Model B) | — | — |
| Exploring the future of privacy-preserving heart disease prediction: a fully homomorphic encryption-driven logistic regression approach HE-aware LR for cardiology |
2025 | CKKS logistic regression outsources heart-disease training to the cloud with only 1% accuracy loss | CKKS | Logistic Regression | Heart Disease (Kaggle) | 84% (−1 pp vs plaintext 85%) | — | — |
| An Efficient Architecture for Edge AI Federated Learning With Homomorphic Encryption Quant+Prune+Cluster FL |
2025 | Federated edge AI with CKKS, quantization, pruning, and clustering boosts accuracy 8.4% and cuts memory 35%. | CKKS | FCNN | Human Activity Recognition (UCI HAR) | 89.86% (+8.4% over FLH baseline) | — | — |
| Enhancing privacy-preserving machine learning with self-learnable activation functions in fully homomorphic encryption Self-learnable polynomial activation |
2024 | Learnable polynomial activation lifts CKKS face-attribute accuracy 0.88-3.15% over square, with sub-second SNN variant | CKKS | CNN (SLAFNN-10 / SNN) | UTKFace | 88.07% gender / 89.16% age / 76.54% ethnicity (SLAFNN-10 ciphertext) | 0.778 | — |
| Practical Considerations of Fully Homomorphic Encryption in Privacy-Preserving Machine Learning Practical FHE benchmarks |
2024 | Hands-on benchmarks show FHE inflates training/inference by 100x-1000x with notable accuracy drops from quantization. | CKKS / BFV (via TenSEAL + Concrete ML) | NN Classifier (Concrete ML) + XGBoost + Logistic Regression | Breast Cancer (Kaggle) | 91.23% XGB / 95.61% NN / 87.72% LR (vs 92.98%/96.49%/92.98% plaintext) | 0.8 | — |
| Private pathological assessment via machine learning and homomorphic encryption FHE SVM kernels |
2024 | CKKS-based SVM inference on encrypted medical tabular and imaging data in seconds at 128-bit security. | CKKS | SVM | Multiple | F1 0.85 (PneumoniaMNIST, RBF); F1 0.96 (WBC, Sigmoid) | 2.03 | GitHub |
| Non-interactive Fully Encrypted Machine Learning Protocol for Inference Non-interactive FE PPML |
2024 | First fully-encrypted FE protocol where the client never comes back online during inference. | FE (DCR-based pkIPFE + LinEnc-QFE) | FCNN (2-layer Quadratic Neural Network) | Iris + Breast Cancer (UCI) | Output error < 1e-7 vs plaintext; classification accuracy not reported | 11952.0 | GitHub |
| Federated Learning with Quantum Computing and Fully Homomorphic Encryption: A Novel Computing Paradigm Shift in Privacy-Preserving ML FHE + Quantum FL |
2024 | CKKS-encrypted federated aggregation of hybrid classical+quantum neural network weights across 20 clients. | CKKS | Hybrid classical+QNN (parameterized quantum circuit) | Multiple | 88.75% Brain MRI; 70.12% CIFAR-10; 70.15% PCOS (test) | — | GitHub |
| Privacy-Preserving Graph-Based Machine Learning with Fully Homomorphic Encryption for Collaborative Anti-Money Laundering TFHE graph-AML pipelines |
2024 | TFHE-encrypted XGBoost with graph features hits >99% accuracy on AML detection; GNN pipeline only partially compatible. | TFHE | GIN (GNN) + XGBoost-GFP | AMLworld HI-Small | 99.72% (FHE = clear) on balanced dataset; F1 0.3867 on imbalanced (+8% vs basic) [§6.4, §6.5] | — | GitHub |
| Secure Federated Learning With Fully Homomorphic Encryption for IoT Communications Cluster-matrix FHE FL |
2024 | Four FHE+FL clustering variants for IoT botnet detection; up to 89.98% comm-overhead and 70.38% latency cuts. | FHE (scheme not specified) | CNN | N-BaIoT | 0.997 avg (OECM/MUCM/MECM); 0.911 (OUCM); baseline 0.996–0.998 [Tables III–VII] | — | GitHub |
| Recent advances of privacy-preserving machine learning based on (Fully) Homomorphic Encryption FHE generations survey |
2024 | Tour of FHE-based PPML across four FHE generations, mixed-mode schemes, and hardware acceleration. | survey (BFV, BGV, CGGI/TFHE, CKKS, mixed) | N/A | Multiple | Not reported | — | — |
| Secure federated learning applied to medical imaging with fully homomorphic encryption FL + CKKS last-layer |
2024 | Federated MobileNetV2 on Belgian mammograms encrypts only the last layer under CKKS due to RAM limits. | CKKS | MobileNetV2 (transfer-learned, binary head) | Private Belgian mammograms | 87.4% with HE vs 73.8% without (batch 32, 3 rounds, 3 clients) | — | — |
| Machine Learning Training on Encrypted Data with TFHE TFHE encrypted training |
2024 | End-to-end encrypted SGD training of LR and small MLPs via TFHE rounding and 4-6 bit quantization. | TFHE | Logistic Regression + MLP (1 hidden) | Multiple | 98.25% (-0.87 vs fp32) breast-cancer LR | — | — |
| Efficiency Optimization Techniques in Privacy-Preserving Federated Learning With Homomorphic Encryption: A Brief Survey HE-PPFL efficiency survey |
2024 | Taxonomy of algorithmic, hardware, and hybrid optimizations for HE-based federated learning on edge devices. | survey (CKKS/BGV/BFV/TFHE/Paillier) | N/A | Multiple | Not reported | — | — |
| HE-MAN – Homomorphically Encrypted MAchine learning with oNnx models ONNX-driven FHE toolkit |
2023 | Auto-tunes Concrete (TFHE) or TenSEAL (CKKS) for any ONNX model with no FHE expertise required. | hybrid (TFHE via Concrete, CKKS via TenSEAL) | CNN (CryptoNets, LeNet-5, MobileFaceNets-classifier) | Multiple | 0.968 CryptoNets / 0.984 LeNet-5 / 0.970 LFW (Concrete, vs 0.975/0.991/0.990 plaintext) | 112.0 | GitHub |
| Optimized Privacy-Preserving CNN Inference With Fully Homomorphic Encryption Constant-time conv FHE |
2023 | Coefficient-packed convolutions make FHE CNN cost constant in kernel size; 398 s for 94% on CIFAR-10. | CKKS | CNN (Plain-20 / Plain-18 variants) | CIFAR-10/100, ImageNet | 94.0% CIFAR-10 / 73.5% CIFAR-100 (state-of-art among FHE prior work) | 398.0 | GitHub |
| Practical Privacy-Preserving Machine Learning using Fully Homomorphic Encryption FHE SVM training |
2023 | Trains a linear SVM on 8,000 encrypted samples in under 45 seconds via levelled CKKS and client-assisted Newton. | CKKS | SVM | Wisconsin Breast Cancer (WBC) | 99.1% (same as plaintext baseline) | — | — |
| Fully Homomorphic Encryption with Applications to Privacy-Preserving Machine Learning Pedagogical FHE thesis |
2023 | Self-contained thesis from LWE foundations to GSW, CKKS, and a multikey-CKKS federated learning protocol. | GSW, CKKS, xMK-CKKS | N/A | Not reported | Not reported | — | — |
| FedML-HE: An Efficient Homomorphic-Encryption-Based Privacy-Preserving Federated Learning System Selective parameter encryption |
2023 | Selective parameter encryption cuts HE-federated overhead ~10x for ResNet-50 and ~40x for BERT. | CKKS | Multiple (Linear, MLP, LeNet, RNN, CNN, MobileNet, ResNet-18/34/50, GroupViT, ViT, BERT, Llama-2 7B) | Multiple | Negligible delta (e.g. 0 to -0.28% on CNN vs plaintext) [Table 6] | — | — |
| FheFL: Fully Homomorphic Encryption Friendly Privacy-Preserving Federated Learning with Byzantine Users Byzantine-robust multi-key FHE |
2023 | Single-server FL that defends against poisoning clients entirely under CKKS — no second server needed. | CKKS (distributed multi-key additive) | CNN / ResNet-18 | Multiple | 98.78% MNIST (-2.7 vs plaintext at N=16384, 100 users) | — | — |
| PPFLHE: A privacy-preserving federated learning scheme with homomorphic encryption for healthcare data Federated Paillier + ACK |
2023 | Paillier-encrypted federated MobileNetV2 for diabetic retinopathy with ACK-based dropout handling. | Paillier (PHE) | MobileNetV2 | APTOS 2019 Blindness Detection | 81.53% (+1.53 vs prior FL [40]) | — | — |
| Secure human action recognition by encrypted neural network inference Fast homomorphic convolution |
2022 | Skeleton-based fall detection under CKKS in 2.4 s with 613x speedup over LoLa. | CKKS | CNN (2D-CNN-128) | J-HMDB + URFD + Multicam | 86.21% sensitivity / 99.14% specificity (−0.16–0.17% vs plaintext) | 2.419 | GitHub |
| Homomorphic Encryption for Machine Learning and Artificial Intelligence Applications Nuclear NDT motivation |
2022 | Argonne proof-of-concept: CKKS-encrypted FCNN on MNIST trades 7.7 pts accuracy for 150,000x slowdown. | CKKS | FCNN | MNIST | 76.1% (-7.7 vs 83.8% plaintext) | 295.3 | — |
| Privacy-Preserving Machine Learning With Fully Homomorphic Encryption for Deep Neural Network First FHE ResNet-20 |
2022 | First standard ResNet-20 on encrypted CIFAR-10 under RNS-CKKS with bootstrapping; 92.43% accuracy in ~3 h. | RNS-CKKS | ResNet-20 | CIFAR-10 | 92.43% +/- 2.65% (vs 91.89% plaintext; 98.43% agreement) | 10800.0 | — |
| Exploring Bitslicing Architectures for Enabling FHE-Assisted Machine Learning Bitsliced TFHE gates |
2022 | Software bitslicing over NuFHE gives 37x speedup for exact-FHE CNN inference in ~700 MB | TFHE (NuFHE) | CNN | MNIST | ~98% on MNIST after INT8 quantization (≤2% drop vs FP32) [§IV-A, Table VI] | 16200.0 | — |
| A Survey of Deep Learning Architectures for Privacy-Preserving Machine Learning With Fully Homomorphic Encryption DL+FHE comparison metrics |
2022 | Cross-cuts FHE-PPML works by NN layer, activation, packing, bootstrapping and proposes hardware-independent comparison metrics. | survey (BFV, BGV, CKKS, TFHE, YASHE) | N/A | Multiple | Not reported | — | — |
| Fully Homomorphically Encrypted Deep Learning as a Service Encrypted sequence model |
2021 | First CKKS encrypted 1D-CNN sequence model for agri-food milk-yield forecasting in a client-server EDLaaS pipeline. | CKKS | 1D CNN | Langhill Dairy herd (30 years) | 87.6% (MAPE 12.4%); MSE/MAE match plaintext to 3 s.f. | 0.966 | GitHub |
| Privacy-preserving neural networks with Homomorphic encryption: Challenges and opportunities NN-HE survey + benchmark |
2021 | Survey of NN-HE with benchmark of CryptoNets and LoLa under SEAL+PyTorch on MNIST/F-MNIST. | BFV (SEAL) + CKKS | CNN | MNIST, F-MNIST | 98.95% MNIST (LoLa-N, SEAL); 99.51% plaintext PyTorch | 3.5 | — |
| Towards the AlexNet Moment for Homomorphic Encryption: HCNN, the First Homomorphic CNN on Encrypted Data With GPUs First GPU HCNN |
2021 | First GPU-accelerated FHE CNN; MNIST in 5.16 s and CIFAR-10 in 304 s under BFV. | BFV | CNN | MNIST + CIFAR-10 | 99% MNIST / 77.55% CIFAR-10 (scalar-encoded) | 5.16 | — |
| DOReN: Toward Efficient Deep Convolutional Neural Networks with Fully Homomorphic Encryption Compressor-based accumulator |
2021 | Exact ReLU on encrypted CNNs via 3:2 compressors, 20x faster amortized than SHE. | BGV | CNN (VGG-7, ResNet-20, LoLa, GHE, SHE) | CIFAR-10 | Not reported | — | — |
| Private Machine Learning Classification Based on Fully Homomorphic Encryption 1-round HE compare |
2020 | Improved BGV scheme plus one-round homomorphic compare speeds private hyperplane, Naive Bayes, and decision-tree inference. | BGV (improved HElib-style) | decision tree / hyperplane / Naive Bayes | Wisconsin Breast Cancer (UCI) | Not reported | 0.000177 | — |
| A Survey on Privacy-Preserving Machine Learning with Fully Homomorphic Encryption FHE+MLaaS survey |
2020 | Survey maps FHE families, bootstrapping, and MLaaS approaches to position privacy-preserving ML directions. | survey (covers lattice, integer, (R)LWE, NTRU) | N/A | N/A (survey) | Not reported | — | — |
| Homomorphic Encryption for Machine Learning in Medicine and Bioinformatics Medical FHE survey |
2020 | Survey of FHE schemes, libraries, and privacy-preserving ML and genomics methods targeted at medical applications. | survey (BGV, BFV, CKKS, TFHE, FHEW) | N/A | Multiple | Not reported | — | — |
| On Fully Homomorphic Encryption for Privacy-Preserving Deep Learning Bitwise bootstrapped CNN |
2019 | Bitwise Gentry-FHE CNN on MNIST exposes bootstrap-per-gate cost — projected ~10^12 s per image at lambda=9. | Gentry bootstrapped FHE (AGCD + SSS) | CNN | MNIST | 91.2% (plaintext float, full-match PR=1.0 at Q12.6 on NI=30) | — | GitHub |
| Fully Homomorphic Encryption for Machine Learning Scale-invariant FHE-DiNN |
2018 | PhD thesis: TFHE per-neuron bootstrapping with sign activation gives depth-independent FHE for discretized neural nets. | TFHE | FCNN (DiNN, 1 hidden layer) | MNIST | 96.35% (-0.08 vs DiNN plaintext, -2.6 vs Cryptonets) | 0.49 | GitHub |
The table is generated from the comparison: and frontpage: YAML blocks at the top of each summaries/*.md file. Do not edit rows by hand — edit the per-paper file and re-run the build (see How to contribute).
Glossary
- CKKS — Approximate homomorphic encryption for fixed-point vectors. Operations are over polynomials with a rescaling step; tolerates small precision loss, so it dominates ML inference work.
- BFV / BGV — Exact integer homomorphic schemes. No approximation, but require careful modulus-chain budgeting for multi-layer networks.
- TFHE — Bit-level homomorphic encryption supporting fast programmable bootstrapping; cheap nonlinearity, expensive linear ops compared to CKKS.
- Bootstrapping — Refreshing a ciphertext to reduce accumulated noise so further homomorphic operations are possible. Expensive; many schemes avoid it via “levelled” parameter sets.
- Polynomial activation — A low-degree polynomial (often degree 2–4) that approximates ReLU / sigmoid / softmax over a training-time input range, since true ReLU is not FHE-friendly.
- SIMD packing — Encoding many input values into one ciphertext slot vector so a single homomorphic op processes them in parallel (used by CKKS/BFV/BGV).
- Hybrid (client/server) FHE — A protocol where the server holds the encrypted state for linear ops, but hands intermediate results to the client to run the nonlinearity in plaintext, then re-encrypts. Avoids bootstrapping at the cost of client interactivity and per-layer ciphertext transfers.
How to contribute
To add a paper:
- Drop the PDF into
papers/(gitignored — local only). - Convert it to text:
pdftotext -layout papers/<filename>.pdf /tmp/<slug>.txt. - Copy
extraction-template.mdtosummaries/<slug>.mdand fill in every section from the text, including thecomparison:andfrontpage:YAML blocks. Follow the slug conventionfirstauthor-year-shortid(lowercase, hyphenated). - Re-run the build script to regenerate the comparison table and stats:
python3 build_index.pyThe script reads every
summaries/*.mdfront matter and rewrites the marker-delimited regions in this file and insummaries/index.md. Do not edit the rows by hand.
See extraction-template.md (per-paper schema) and frontpage-template.md (front-page schema, controlled vocabularies, anti-patterns).