PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning a Fixed-Length Fingerprint Representation

Joshua J. Engelsma, Kai Cao, Anil K. Jain

arXiv:1909.09901v2cs.CV

TL;DR

Variable-length minutiae representations pose challenges for fingerprint comparison, and reliable minutiae extraction can be difficult in low-quality images. DeepPrint learns a compact fixed-length representation using alignment and minutiae detection as domain guidance. On a 1.1-million-fingerprint gallery, it achieves competitive search accuracy at substantially faster speeds than commercial matchers.

  • Problem

    Variable-length minutiae representations remain prevalent, while reliable minutiae extraction is problematic in low-quality fingerprints.

    Source: §1 · p. 1

  • Method

    DeepPrint learns a 200-byte fixed-length fingerprint representation using alignment and minutiae detection to guide its network.

    Source: §1 · pp. 2–3

  • Results

    300 ms vs. 27,000 ms against a gallery of 1.1 million: DeepPrint reports competitive search accuracy at significantly faster speeds than commercial matchers.

    Source: §1 · p. 4, §14 · p. 15

  • Takeaways & Limitations

    DeepPrint provides fixed-length representations for authentication and search, and the paper demonstrates encrypted-domain matching using fully homomorphic encryption.

    Source: §14 · pp. 14–15

    Search performance on FVC 2004 DB1 A is not reported because its 700 slap prints do not provide a sufficiently large background for meaningful search results.

    Source: §9.1 · p. 12

Abstract

from arXiv · show

We present DeepPrint, a deep network, which learns to extract fixed-length fingerprint representations of only 200 bytes. DeepPrint incorporates fingerprint domain knowledge, including alignment and minutiae detection, into the deep network architecture to maximize the discriminative power of its representation. The compact, DeepPrint representation has several advantages over the prevailing variable length minutiae representation which (i) requires computationally expensive graph matching techniques, (ii) is difficult to secure using strong encryption schemes (e.g. homomorphic encryption), and (iii) has low discriminative power in poor quality fingerprints where minutiae extraction is unreliable. We benchmark DeepPrint against two top performing COTS SDKs (Verifinger and Innovatrics) from the NIST and FVC evaluations. Coupled with a re-ranking scheme, the DeepPrint rank-1 search accuracy on the NIST SD4 dataset against a gallery of 1.1 million fingerprints is comparable to the top COTS matcher, but it is significantly faster (DeepPrint: 98.80% in 0.3 seconds vs. COTS A: 98.85% in 27 seconds). To the best of our knowledge, the DeepPrint representation is the most compact and discriminative fixed-length fingerprint representation reported in the academic literature.

1 INTRODUCTION

DeepPrint replaces variable-length minutiae sets with a compact, fixed-length representation that combines fingerprint-domain knowledge with learned features. The paper reports competitive search accuracy against commercial matchers at much faster speeds, alongside applications to encrypted matching and candidate reranking.

1 INTRODUCTION

  • The paper motivates fixed-length representations by noting that minutiae counts vary and reliable extraction can be problematic in low-quality fingerprints.The cited examples include false rejects associated with distortion and finger condition.Source: §1 · p. 1, Figure 2 · p. 2
  • DeepPrint learns a 200-byte fixed-length representation using automatic alignment and a minutiae-detection side task alongside texture- and minutiae-based feature branches.The two 96-feature branches are concatenated; minutiae need not be stored in the final template.Source: §1 · pp. 2–3, Figure 4 · p. 3
  • Fixed-length representations can use simple distance metrics and support homomorphic-encryption matching, unlike variable-length minutiae representations that require computationally expensive graph matching.The paper also notes that DeepPrint can encode features beyond minutiae, helping match poor-quality fingerprints when reliable minutiae extraction is unavailable.Source: §1 · p. 2
  • An optional minutiae-map reranking stage can improve accuracy and interpretability, and the authors describe DeepPrint features as complementary to traditional minutiae representations.The map can be converted to a minutiae set and passed to a minutiae matcher.Source: §1 · pp. 3–4
  • Product Quantization compresses gallery representations for mega-scale search, with a reported 51 ms search time against 1.1 million fingerprints versus 27,000 ms for a COTS matcher of comparable accuracy.The paper also notes that Product Quantization can compress representations to 64 bytes with minor accuracy loss.Source: §1 · pp. 3–4
  • DeepPrint offers competitive accuracy and significantly faster search than two state-of-the-art commercial matchers on a gallery of 1.1 million fingerprints.The comparison benchmarks Innovatrics and Verifinger on NIST SD4 and SD14.Source: §1 · pp. 3–4
  • DeepPrint shows comparable authentication performance across rolled NIST SD4 and SD14 fingerprints and slap FVC 2004 DB1 A fingerprints.The paper also reports real-time homomorphic-encryption matching at 1.26 ms with minimal accuracy loss.Source: §1 · p. 4

2 PRIOR WORK

Prior work developed fixed-length fingerprint representations, including newer deep-network approaches, but had not established competitive recognition accuracy or extensive large-scale evaluation across these methods.

2 PRIOR WORK

  • Earlier fixed-length approaches used Gabor-filter global features or minutiae-based descriptors, while recent deep methods aggregate global, multi-resolution, or local descriptors.The cited deep approaches include global alignment, patch-level inference, and aggregation networks.Source: §2 · p. 4
  • Prior deep representations faced trade-offs in alignment cost, pipeline complexity, slap-fingerprint evaluation, occlusion handling, and interpretability.The cited review notes that accuracy for was not reported on slap fingerprints and that averaging local descriptors may be ill-suited to missing or occluded regions.Source: §2 · pp. 4–5
  • Large-scale evaluation remained limited: aside from Cao and Jain testing 250K fingerprints, the largest gallery among the cited studies was 2,700.The paper identifies large-scale search as a primary motivation for fixed-length representations.Source: §2 · p. 5
  • Deep learning had also improved individual fingerprint-system modules, but those works still matched unordered, variable-length minutiae sets.Source: §2 · p. 5

3 DEEPPRINT

DeepPrint combines learned fingerprint representations with automatic alignment and minutiae-map supervision. Its architecture pairs texture and minutiae-guided identity branches, then compresses their combined representation to a 200-byte template.

3.1 Overview

  • DeepPrint trains on 455K longitudinal impressions from 38,291 fingers to learn representations robust to acquisition differences over time and across environments.The main training task predicts finger identity.Source: §3.1 · p. 5

3.2 Alignment

  • A shallow localization network and grid sampler align each 448 × 448 input before the Inception v4 stem and representation branches process it.The localization network’s transformation is optimized through the base-network and representation-extraction losses rather than reference-point supervision.Source: §3.1 · p. 5, §3.2 · p. 6

3.3 Minutiae Map Domain Knowledge

  • The minutiae map is a six-channel heatmap encoding minutiae locations and orientations, used as training supervision rather than as part of the final fixed-length representation.Each map location sums spatial and orientation contributions from minutiae; the map can also be converted back to a minutiae set.Source: §3.3 · p. 7

3.4 Multi-Task Architecture

  • Shared parameters couple identity representation learning to minutiae-map estimation, while a separate texture branch learns complementary identity cues.The texture and minutiae representations each have 96 features; the minutiae map is the secondary task for the minutiae branch.Source: §3.4 · pp. 7–8
  • The network combines identity-classification and center-loss terms with a mean-squared-error minutiae-map loss during multitask training.The authors report 140K training steps over 25 hours, with data augmentation and empirically set loss weights.Source: §3.4 · p. 8

3.5 Template Compression

  • Min-max normalization converts the 192-dimensional, 768-byte floating-point representation into a 200-byte template: 192 feature bytes plus two 4-byte decompression parameters.The paper states that compression only minimally impacts matching accuracy.Source: §3.5 · pp. 8–9, Table 4 · p. 9

4 DEEPPRINT MATCHING

DeepPrint compares unit-length representations with cosine similarity and can fuse its scores with existing minutiae matchers. The paper reports low-cost matching and improved authentication accuracy from fusion.

4 DEEPPRINT MATCHING

  • 192 multiplications and 191 additions suffice for DeepPrint 1:1 authentication using cosine similarity.The authors found cosine similarity consistently outperformed Euclidean distance; with compressed features, authentication time effectively doubles for decompression.Source: §4 · p. 9

4.1 Fusion of DeepPrint Score with Minutiae Score

  • Fusing DeepPrint scores with minutiae-matcher scores improved authentication accuracy, indicating complementary information with little extra computational overhead.DeepPrint’s extracted minutiae can be passed directly to the minutiae matcher, avoiding a second extraction.Source: §4.1 · p. 9

5 DEEPPRINT SEARCH

DeepPrint enables exhaustive gallery search with fixed-length representations, while product quantization reduces search cost and minutiae-based reranking can improve accuracy.

5 DEEPPRINT SEARCH

  • DeepPrint searches all gallery templates using similarity comparisons with complexity linear in gallery size N and representation dimensionality d.The representation has d = 192 dimensions.Source: §5 · p. 9

5.1 Faster Search

  • Product quantization compresses representations into codeword indices, allowing comparisons to use lookup-table operations while retaining exhaustive search.The codebooks are computed offline using k-means clustering.Source: §5.1 · pp. 9–10
  • For 100 million gallery entries, a priority queue of size k = 100 reduced sorting time from 23 seconds to 8 seconds.The 100-million-entry search time was simulated using random representations.Source: §5.1 · p. 9
  • With z = 256 and m = 64, product quantization compresses each gallery template to 64 bytes and replaces 192 additions and multiplications with 64 lookups and additions.The method uses an m × z lookup table built once for matching quantized gallery templates.Source: §5.1 · p. 10

5.2 Two-stage DeepPrint Search

  • DeepPrint retrieves top-k candidates, then reranks them using fused minutiae-matcher and DeepPrint similarity scores.DeepPrint’s minutiae maps are converted to minutiae sets for the minutiae matcher.Source: §5.2 · p. 10

6 SECURE DEEPPRINT MATCHING

DeepPrint’s fixed-length representation supports matching under fully homomorphic encryption because its similarity computation uses additions and multiplications. The reported authentication performance remains almost unchanged, unlike minutiae-based matching, which cannot be reduced to those operations.

6 SECURE DEEPPRINT MATCHING

  • DeepPrint similarity uses only multiplication and addition, whereas minutiae matching cannot be reduced to those operations for fully homomorphic encryption.The cited fuzzy-vault schemes for minutiae templates lose matching accuracy and are sensitive to pre-alignment.Source: §6 · p. 10
  • DeepPrint authentication performance remains almost unaltered under fully homomorphic encryption, with precision loss from converting floating-point features to integers.The representation can be matched with 192 bits of security.Source: §6 · p. 10

7 DATASETS

The experiments use rolled-fingerprint training data and test on rolled and slap datasets. The benchmark discussion notes that small gallery results are saturated and motivates comparison against a 1.1-million-fingerprint background.

7 DATASETS

  • Training uses 455K rolled fingerprint images from 38,291 unique fingers, while testing includes rolled NIST SD4 and SD14 images and slap images from FVC 2004 DB1 A.The NIST SD14 evaluation uses its last 2,700 pairs for comparison with previous approaches.Source: §7 · p. 10, §7.1 · p. 10

7.1 NIST SD4 & NIST SD14

  • NIST SD4 contains 2,000 unique rolled-fingerprint pairs, and NIST SD14 contains 27,000 pairs; evaluation uses the last 2,700 SD14 pairs.SD4 pairs are evenly distributed across five fingerprint types.Source: §7.1 · p. 10
  • The benchmark results show DeepPrint competitive with the best fixed-length representation and state-of-the-art COTS, while small-gallery performance is saturated.The paper therefore compares against state-of-the-art COTS using a background of 1.1 million fingerprints.Source: §7.1 · p. 11
  • FVC 2004 DB1 A tests generalization from rolled-image training to slap fingerprints despite small overlap, heavy nonlinear distortion, and wet or dry fingers.The authors attribute this generalization to incorporating domain knowledge into the network architecture.Source: §7.2 · p. 11

8 COTS MATCHERS

The paper benchmarks DeepPrint against two commercial fingerprint SDKs, using their more discriminative proprietary templates.

8 COTS MATCHERS

  • The benchmarks use COTS A and COTS B, identified as Verifinger 10.0 or Innovatrics v7.2.1.40, the latest SDK versions as of July 2019.The nondisclosure agreement prevents linking either alias to a specific SDK.Source: §8 · p. 11
  • Both SDKs provide ISO minutia-only and proprietary templates, and the study uses the more discriminative proprietary templates for best performance.Those proprietary templates contain minutiae and other features unknown to the authors.Source: §8 · p. 11
  • Verifinger and Innovatrics are described as top performers in NIST and FVC evaluations.Source: §8 · p. 11

9 BENCHMARK EVALUATIONS

The benchmarks compare DeepPrint with fixed-length search representations and commercial minutiae matchers, then assess authentication, score fusion, and encrypted authentication. DeepPrint is competitive on search, outperforms COTS B across authentication protocols, and retains nearly the same authentication accuracy after encryption.

9.1 Search (1:N Comparison)

  • DeepPrint is competitive with the most accurate previously published search algorithm [28], with slightly lower performance on NIST SD4 and slightly higher performance on NIST SD14.The authors note that existing NIST search benchmarks are saturated and use a 1.1-million-fingerprint gallery in later experiments.Source: §9.1 · p. 12

9.2 Authentication

  • DeepPrint outperforms COTS B on all benchmark authentication protocols and both COTS A and COTS B on FVC 2004 DB1 A.The network was trained on rolled fingerprints, while FVC 2004 DB1 A contains slap fingerprints with different textural characteristics.Source: §9.2 · p. 12, Table 6 · p. 12, Table 7 · p. 12

9.2.1 Fusion with Minutiae-Matchers

  • Score-level fusion improves authentication performance, consistent with DeepPrint’s peaked genuine-score distribution complementing minutiae matchers’ strong true-reject behavior.A DeepPrint score requires 192 multiplications and 191 additions, adding little overhead for existing COTS matchers.Source: §9.2.1 · p. 12

9.2.2 Secure Authentication

  • After fully homomorphic encryption, DeepPrint authentication accuracy remains nearly the same, and matching a pair of templates takes 1.26 milliseconds.The reported encrypted-domain authentication result appears in Table 8.Source: §9.2.2 · p. 12

10 LARGE SCALE SEARCH

DeepPrint search is evaluated on galleries of more than 1.1 million fingerprints using exhaustive matching, minutiae re-ranking, and product quantization. Re-ranking yields accuracy competitive with the top COTS SDK at substantially faster speeds, while quantization further reduces search time and template size.

10.1 DeepPrint Search

  • 160 milliseconds yields Rank-1 identification accuracies of 95.15% on NIST SD4 and 94.44% on NIST SD14 with exhaustive search.At Rank-100, both datasets exceed 99% accuracy.Source: §10.1 · p. 13, Figure 10 · p. 13

10.3 Product Quantization

  • 51 milliseconds after product quantization, down from 160 milliseconds, with Rank-1 accuracies of 94.8% on NIST SD4 and 94.2% on NIST SD14.Product quantization reduces the DeepPrint template to 64 bytes; before quantization, the respective Rank-1 accuracies were 95.15% and 94.44%.Source: Figure 10 · p. 13, §10.3 · p. 13, Table 10 · p. 13

10.2 Minutiae Re-ranking

  • Re-ranking DeepPrint’s retrieved candidates achieves search accuracy competitive with the top COTS SDK at significantly faster speeds.The COTS matcher re-ranks the top 500 DeepPrint candidates; Table 9 reports the comparison on a 1.1-million-fingerprint background.Source: §10.2 · p. 13, Table 9 · p. 13

11 ABLATION STUDY

The ablation study finds that domain knowledge improves authentication performance, with alignment benefits depending on the fingerprint dataset. It also reports complementary minutiae and texture representations whose score fusion improves accuracy.

11 ABLATION STUDY

  • Adding domain knowledge improves authentication performance in all reported scenarios.The study reports this result in Tables 11 and 12.Source: §11 · p. 13
  • Alignment does not help on already mostly aligned NIST SD4 and SD14 fingerprints, but significantly improves performance on likely severely unaligned FVC 2004 DB1 A images.FVC slap fingerprints differ from the rolled fingerprints used for training in size, distortion, and conditions.Source: §11 · pp. 13–14
  • Fusing minutiae-based and texture-based representation scores improves accuracy, supporting their complementarity.The figure description shows the texture branch focusing around singularity points and the minutiae branch around minutiae points.Source: §11 · p. 14, Figure 11 · p. 14

12 INTERPRETABILITY

DeepPrint’s texture and minutiae branches focus on complementary fingerprint regions, and their fused representations improve matching performance. The conclusion presents a 200-byte representation for authentication and search, with fast large-gallery search and encrypted-domain matching.

12 INTERPRETABILITY

  • The texture branch focuses around singularity points, while the minutiae branch attends to regions with high minutiae density; fusing their representations improves matching performance.The authors use a deconvolutional network to visualize the pixels attended to by each branch.Source: §12 · p. 14

13 COMPUTATIONAL RESOURCES

  • DeepPrint models use TensorFlow 1.14.0 and were trained on two NVIDIA GeForce RTX 2080 Ti GPUs; experiments ran on an Intel Core i9-7900X CPU with 32 GB RAM.Source: §13 · p. 14

14 CONCLUSION

  • The conclusion reports competitive search accuracy at significantly faster speeds on a 1.1-million-fingerprint gallery and matching through fully homomorphic encryption.DeepPrint extracts a fixed-length 200-byte representation for both authentication and search.Source: §14 · pp. 14–15
Loading 1909.09901v2…