Research map / Federated learning and privacy / Privacy attacks and secure learning
Blind-Touch: Homomorphic Encryption-Based Distributed Neural Network Inference for Privacy-Preserving Fingerprint Authentication
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Blind-Touch runs a Siamese CNN fingerprint matcher in a distributed setup where the client computes FC-16 in plaintext and the server evaluates FC-1 over CKKS-encrypted features, with a compression step and clustered servers for scalable 1:N matching.
Problem
Fingerprint authentication for web and cloud services is limited by privacy risks of storing and processing biometric data on servers, and existing HE-based approaches are too slow for practical 1:N search over thousands of users.
Approach
Split the network so the client computes the first FC layer in plaintext (a linear map that commutes with subtraction under encryption), the server runs only the subtraction, square activation and FC-1 under CKKS, a one-hot compression packs up to 8,192 results into one ciphertext, and multiple cluster servers search in parallel.
Key results
- F1-score of 93.8% on PolyU (93.6% at threshold 0.1) and 98.2% F1 on SOKOTO at threshold 0.01.
- Authentication among 5,000 fingerprints in about 650 ms with three clusters; total time 511 ms summed across tasks.
- Input vector about 68x smaller than DeepPrint w/ BFV and authentication about 9x faster; PolyU AUC 97.5% and EER 2.5%, close to plaintext SOTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compute FC-16 after encryption rather than in plaintext on the client (ablation) | Blind-Touch · 5,000 fingerprints | Evaluation time: 0.65 seconds -> 15,096 seconds | Hurt |
| remove the authentication-result compression method (ablation) | Blind-Touch · 5,000 inputs | Ciphertext size: 0.26 MB -> 2.6 MB | Hurt |
| use 1 cluster instead of 3 (ablation) | Blind-Touch · 5,000 registered fingerprints | Authentication time: 650.2 -> 1,334.4 | Hurt |
| use 2 clusters instead of 3 (ablation) | Blind-Touch · 5,000 registered fingerprints | Authentication time: 650.2 -> 790.5 | Hurt |
| increase feature vector size from 16 to 64 (ablation) | Blind-Touch · 5,000 fingerprints | Total time: 0.65 seconds -> 1.81 seconds | Hurt |
| use a 4-layer CNN instead of a 5-layer CNN (ablation) | Blind-Touch · PolyU | F1-score: 93.8% -> 92.4% | Hurt |
| use a 6-layer CNN instead of a 5-layer CNN (ablation) | Blind-Touch · PolyU | F1-score: 93.8% -> 89.3% | Hurt |
| use Blind-Touch rather than DeepPrint w/ CKKS | PolyU authentication · 5,000 fingerprints | Authentication time: 772 -> 485 | Helped |
| use Blind-Touch rather than DeepPrint w/ BFV | PolyU authentication · 5,000 fingerprints | Authentication time: 4,297 -> 485 | Helped |
| use Blind-Touch rather than ContactlessMinuNet | PolyU fingerprint authentication · PolyU | AUC: 99.3 -> 97.5 | Hurt |
| use Blind-Touch rather than ContactlessMinuNet | PolyU fingerprint authentication · PolyU | EER: 1.9 -> 2.5 | Hurt |
Surprising findings
- Decryption is slower for Blind-Touch than for DeepPrint (8 ms vs 1 ms for BFV), a noted drawback.
- Moving the sigmoid to the client keeps the server-side HE depth to three multiplications.
Limitations
- Security proof is limited to a semi-honest curious server under IND-CPA; no malicious-model analysis.
- Decryption time is higher than DeepPrint baselines.
- Rank-1 accuracy of 91% reported only on the distorted SOKOTO set.
Experimental setup
Datasets: PolyU Cross Sensor Fingerprint Database SOKOTO Coventry fingerprint dataset
Benchmarks: PolyU Cross Sensor Fingerprint Database SOKOTO Coventry dataset
Metrics: accuracy F1-score AUC EER Rank-1 accuracy execution time ciphertext size
Baselines: DeepPrint w/ CKKS DeepPrint w/ BFV MNIST mindtct MinutiaeNet VeriFinger ContactlessMinuNet MinNet
Builds on: CKKS DeepPrint CryptoNets Siamese neural network
Closest papers by meaning
- Intersecting-Boundary-Sensitive Fingerprinting for Tampering Detection of DNN Models (ICML 2024 · similarity 0.82)
- CryptoFace: End-to-End Encrypted Face Recognition (CVPR 2025 · similarity 0.82)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.82)
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.82)
- OmniMark: Efficient and Scalable Latent Diffusion Model Fingerprinting (AAAI 2025 · similarity 0.82)
- SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.