A privacy preserving machine learning framework for medical image analysis using quantized fully connected neural networks with TFHE based inference¶
Authors: Sadhana Selvakumar; B. Senthilkumar
Year: 2025
Venue: Scientific Reports
type: application collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
The paper builds a privacy-preserving medical image classification pipeline in which a fully connected neural network is trained in plaintext with quantization-aware training, validated under FHE constraints, and compiled with Concrete ML into a TFHE circuit for encrypted inference on MedMNIST subsets. The abstract reports 88.2% plaintext versus 87.5% encrypted accuracy with about 150 ms average inference per image, while the conclusion reports per-dataset encrypted accuracies of 88.88% (PneumoniaMNIST), 81.81% (BreastMNIST) and 84.61% (BloodMNIST) and a per-sample inference time rising from 0.31 ms in plaintext to 119.55 s under TFHE. The authors conclude that accuracy loss is marginal but the computational overhead is substantial.
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | PPML framework for medical image analysis with quantized FCNN and TFHE inference | llm | 0.65 |
| contribution | quantization-aware training of the FCNN for FHE compatibility | llm | 0.65 |
| contribution | FHE simulation and compilation of the quantized model with Concrete ML | llm | 0.65 |
| contribution | evaluation of plaintext vs encrypted accuracy and inference time on MedMNIST subsets | llm | 0.65 |
| schemes | TFHE | llm | 0.65 |
| libraries | Concrete ML | llm | 0.65 |
| models | quantized fully connected neural network (FCNN) | llm | 0.65 |
| datasets | MedMNIST | llm | 0.65 |
| datasets | PneumoniaMNIST | llm | 0.65 |
| datasets | BreastMNIST | llm | 0.65 |
| datasets | BloodMNIST | llm | 0.65 |
| tasks | medical image classification | llm | 0.65 |
| tasks | private inference | llm | 0.65 |
| threat model | untrusted third-party inference service | llm | 0.65 |
| interaction | non-interactive FHE | llm | 0.65 |
| code url | https://github.com/zama-ai/concrete-ml (URL found in the full text) | regex | 0.50 |
| limitations | scope of experimentation limited to selected MedMNIST subsets; generalisation not comprehensively assessed | llm | 0.65 |
| limitations | significant computational overhead of TFHE inference (119.55 s vs 0.31 ms per sample on PneumoniaMNIST) | llm | 0.65 |
| llm relevance | medium | llm | 0.65 |
| key result | accuracy (plaintext): 88.2% (FCNN on MedMNIST, abstract) | llm | 0.65 |
| key result | accuracy (encrypted): 87.5% (TFHE inference on MedMNIST, abstract) | llm | 0.65 |
| key result | average inference time: 150 ms per image (encrypted inference, abstract) | llm | 0.65 |
| key result | accuracy (encrypted, per dataset): 88.88% / 81.81% / 84.61% (PneumoniaMNIST / BreastMNIST / BloodMNIST, conclusion) | llm | 0.65 |
| key result | inference time per sample (encrypted): 119.55 s (PneumoniaMNIST, TFHE, conclusion) | llm | 0.65 |
| key result | inference time per sample (plaintext): 0.31 ms (PneumoniaMNIST, conclusion) | llm | 0.65 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: TFHE 74 [p.1], CKKS 20 [p.3], BFV 13 [p.3], MPC 7 [p.2], BGV 4 [p.3], FHEW 3 [p.5], Differential privacy 3 [p.2], Leveled / somewhat HE 3 [p.4]
- library: Concrete (Zama) 8 [p.7], Concrete ML 8 [p.7], Zama 1 [p.22]
- model: CNN 16 [p.1], SVM 1 [p.4]
- dataset: MedMNIST 37 [p.1], Chest X-ray 1 [p.15]
- technique: Quantization 57 [p.1], Bootstrapping 31 [p.2], Programmable bootstrapping 24 [p.2], Pruning 18 [p.2], Federated learning 14 [p.2], ReLU 5 [p.6], Convolution 4 [p.1], Fine-tuning 3 [p.7], Functional bootstrapping 2 [p.4], Encrypted training 2 [p.18]
- assumption: LWE 12 [p.6], IND-CPA 8 [p.21], Lattice-based 2 [p.5]
- task: Medical / healthcare 128 [p.1], Image classification 3 [p.8], Finance 3 [p.4]
- topic: Homomorphic encryption 141 [p.1], Privacy-preserving machine learning 46 [p.1], Private / secure inference 29 [p.1], Benchmark 2 [p.15], Open source 2 [p.7]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/applications/selvakumar-2025_ppml-framework-for-medical-image-analysis-using-quantized-fully-connected-networks.pdf |
ppml/applications | 23 | 3.8 MB | d1956df4658e |
was Selvakumar and Senthilkumar - 2025 - A privacy preserving machine learning framework for medical image analysis using quantized fully con.pdf |
Database key: a-privacy-preserving-machine-learning-framework-for-medical-image · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set a-privacy-preserving-machine-learning-framework-for-medical-image COLUMN VALUE