TL;DR

A privacy-preserving fingerprint authentication pipeline extracts CNN feature vectors in plaintext on the client, encrypts them under CKKS via TenSEAL, and performs encrypted squared-Euclidean matching against stored encrypted templates, achieving 99.06% accuracy with EER 0.40% and a 0.025 s encrypted comparison on a SOCOFing subset of 90 users [Abstract][§IX].

Problem and motivation

Fingerprints are immutable and sensitive: stolen biometric templates can be inverted to reconstruct images or used for impersonation, and GDPR / ISO/IEC 24745 require irreversibility, unlinkability and renewability for biometric template protection [§I.A, §I.B]. Traditional cancelable biometrics and cryptobiometrics trade accuracy for privacy and often rely on auxiliary data [§I.B]. The work targets cloud-based fingerprint authentication for healthcare, banking and access control, where the server must never see raw fingerprint data [§I, §I.E]. Threat model: honest-but-curious authentication and database servers that follow the protocol but may try to infer information from encrypted data, and assume the two servers do not collude [§III.E, §IV.G].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
SOCOFing (subset of 90 individuals) Fingerprint verification 7,200 altered images split 80/10/10 (train/val/test); 1 real fingerprint per user enrolled as encrypted template (90 templates) Fingerprint images, 500 dpi STRANGE-toolbox alterations at three difficulty levels (easy/medium/hard); min-max normalized to [0,1] then scaled by 10^3 before encryption [§IV.A, §IV.B, §IV.C, §III.C].

Pipeline diagram

flowchart LR
    subgraph Client
        A[Capture fingerprint] --> B[CNN feature extract]
        B --> C[Normalize and scale]
        C --> D[Encrypt with pk]
    end
    subgraph DB_Server [Database Server]
        E[(Encrypted templates)]
    end
    subgraph Auth_Server [Authentication Server]
        G[Decrypt score]
        G --> H[Threshold check]
    end
    D -. enc feature .-> Client
    E -. enc template .-> Client
    Client -- enc Euclidean score --> Auth_Server
    H -- accept/deny --> Client

Pipeline steps (text)

  1. Client captures fingerprint and runs the CNN locally to obtain a plaintext feature vector FU [§III.D, §IV.D].
  2. Client normalizes FU to [0,1] then scales to integer range [0, 10^3] for CKKS numerical stability [§III.C].
  3. Client encrypts FU with the public key using TenSEAL CKKS, producing E(FU) [§III.B, §IV.J.1].
  4. Database server returns the corresponding encrypted reference template E(TD) to the client [§III.D, §IV.J.1].
  5. Client computes the encrypted squared Euclidean distance E(Seuc) homomorphically between E(FU) and E(TD) [§III.C.1, §IV.F].
  6. Client sends the encrypted similarity score to the authentication server [§III.D, §IV.F].
  7. Authentication server decrypts the score with the secret key, compares against threshold δ=9, and returns accept/deny [§III.D, §IV.F, §VI.A].

Architecture diagram

flowchart LR
    I([Fingerprint image]) --> C1[Conv + BN]
    C1 --> P1[MaxPool]
    P1 --> C2[Conv + BN]
    C2 --> P2[MaxPool]
    P2 --> D1[Dropout]
    D1 --> F[Flatten]
    F --> FC1[Dense]
    FC1 --> FC2[Dense]
    FC2 --> O([Softmax / feature vector])

Note: exact layer counts, kernel sizes and channel widths are summarized in Table 2 of the paper but are not present in the extracted text [§IV.G, Table 2].

Results

Metric This paper Baseline Hardware
Test accuracy 99.06% [§VI.A, §IX.B] “nearly identical” plaintext accuracy [§VI.A] Dell R750XA, 2x A100 80GB, dual Xeon Silver 4314 [§IV.F]
EER 0.40% (threshold 9, Euclidean) [§VI.A, §VI.D] 0.20% plaintext Euclidean [§VI.D, Table 4] same
TAR 99.19% [§VI.A] — same
FAR 0% [§VI.A, §IX.B] — same
FRR 0.81% (encrypted 0.0081 vs 0.0080 plaintext) [§VI.D, §IX.B] — same
TRR 100% [§IX.B] — same
Test loss 0.1692 [Abstract, §X] — same
Single encrypted comparison 0.025 s (Euclidean) [§IX.D] 0.000162 s plaintext [§VIII.D] same
Total processing per fingerprint 0.136 s [Abstract, §IX.D] — same
Encryption overhead vs plaintext 157.26x [Abstract, §VIII.D] — same
Encrypted vs plaintext matching 90x slower [§IX.E] — same
Storage per template 344 KB; 31 MB for 90 users [Abstract, §IX.D] — same
1M comparisons (identification) ~8 minutes with parallelization [§VIII.D] — same
Cosine alternative 0.028 s encrypted vs 0.00018 s plaintext (156x slower) [§VI.D] — same

Limitations and assumptions

The paper does not provide a like-for-like accuracy / latency comparison table against these systems.

Code and artifacts

Not released. No repository URL or licence is mentioned in the extracted text.

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client - holds sk] -- enc feature E(FU) --> A[Auth Server]
    D[DB Server - stores E(TD)] -- enc template E(TD) --> C
    C -- enc score E(S) --> A
    A -- accept / deny --> C

The client is the only entity holding both raw biometrics and the secret key. Database server only ever sees encrypted templates; authentication server only ever sees the encrypted similarity score (and decrypts it) [§III.D, §III.E, §IV.G].

Open questions