PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Representations for Iris, Face, and Fingerprint Spoofing Detection

David Menotti, Giovani Chiachia, Allan Pinto, William Robson Schwartz, Helio Pedrini, Alexandre Xavier Falcao, Anderson Rocha

arXiv:1410.1980v3cs.CV

TL;DR

Anti-spoofing methods are often tied to their target modality and rely on expert knowledge for feature engineering. This work investigates two deep-representation approaches and reports state-of-the-art performance in eight of nine benchmarks.

  • Problem

    Anti-spoofing methods are usually connected to the modality for which they were designed and often rely on expert knowledge to engineer features.

  • Method

    The work investigates two deep representation approaches and examines their interplay, incorporating experience to develop the SpoofNet network.

  • Results

    Eight of nine benchmarks achieved results outperforming state-of-the-art counterparts.

  • Takeaways & Limitations

    The approaches achieved outstanding classification results across the problems and modalities, with one approach outperforming state-of-the-art results in almost all benchmarks.

Abstract

from arXiv · show

Biometrics systems have significantly improved person identification and authentication, playing an important role in personal, national, and global security. However, these systems might be deceived (or "spoofed") and, despite the recent advances in spoofing detection, current solutions often rely on domain knowledge, specific biometric reading systems, and attack types. We assume a very limited knowledge about biometric spoofing at the sensor to derive outstanding spoofing detection systems for iris, face, and fingerprint modalities based on two deep learning approaches. The first approach consists of learning suitable convolutional network architectures for each domain, while the second approach focuses on learning the weights of the network via back-propagation. We consider nine biometric spoofing benchmarks --- each one containing real and fake samples of a given biometric modality and attack type --- and learn deep representations for each benchmark by combining and contrasting the two learning approaches. This strategy not only provides better comprehension of how these approaches interplay, but also creates systems that exceed the best known results in eight out of the nine benchmarks. The results strongly indicate that spoofing detection systems based on convolutional networks can be robust to attacks already known and possibly adapted, with little effort, to image-based attacks that are yet to come.

I. INTRODUCTION

The paper develops general-purpose convolutional-network anti-spoofing systems for iris, face, and fingerprint modalities with limited sensor-domain knowledge. It contrasts architecture and filter optimization across nine benchmarks, exceeding state-of-the-art results in eight cases.

  • Motivation: Biometric systems support identification, authentication, access control, surveillance, and security, but can be deceived through direct spoofing attacks.Attackers can imitate biometric readings using common apparatus and consumer electronics.
  • Motivation: Existing anti-spoofing systems often depend on modality-specific expert features and attack telltales, so small attack changes may require redesigning the entire system.The paper avoids custom-tailored solutions and assumes little prior knowledge about the problem.
  • Approach: The study explores architecture optimization and filter optimization to build convolutional anti-spoofing systems across iris, face, and fingerprint attack types.Architecture optimization searches network designs with random filter weights, whereas filter optimization learns weights by back-propagation in a predefined architecture.
  • Approach: Architecture and filter optimization are evaluated separately and in combination to study their interplay and construct systems for nine public benchmarks.The approaches are crossed by applying learned architectures with random filters and architecture-specific filters learned by back-propagation.
  • Results: Eight of nine benchmark results outperform state-of-the-art counterparts, while the better approach varies with sample size and problem difficulty.Combining both approaches sometimes achieves performance unavailable from either approach alone.
  • Results: The experiments strongly indicate that convolutional networks can provide robust spoofing detection and may adapt with little effort to future image-based attacks.The study also proposes a new convolutional architecture that improves performance in five problems.

II. RELATED WORK

Prior iris anti-spoofing research uses hardware, image-quality measures, texture patterns, and frequency or noise artifacts. This paper instead learns vision-relevant features directly from data using deep representations.

  • Iris spoofing: Iris liveness methods range from special acquisition hardware to software using image quality, texture, spectral, and recapture-noise cues.Examples include focus, motion, occlusion, pupil dilation, contact-lens effects, and high-frequency spectral magnitude.
  • Iris spoofing: Prior iris methods validate classifiers and selected features on benchmarks including BioSec, Clarkson, MobBIOfake, Warsaw, and NotreDame.Reported approaches include sequential feature selection, iris segmentation, and hierarchical visual codebooks.
  • Iris spoofing: Performance of existing iris anti-spoofing solutions varies significantly from dataset to dataset.The paper presents deep representations as an alternative to hard-coded feature engineering.

B. Face Spoofing

Related spoofing-detection work spans face and fingerprint modalities, relying largely on texture, quality, hardware, and data-driven descriptors. Fingerprint spoofing remains an open problem with results far from perfect.

  • Face spoofing: Face anti-spoofing methods include user-behavior modeling, extra devices, user cooperation, texture analysis, reflectance, and data-driven approaches.These methods address photo, video, and 3D-mask attacks.
  • Face spoofing: Face research uses temporal rhythm, dynamic texture, LBP-like descriptors, Gabor methods, reflectance, and score-level fusion for spoofing detection.Dynamic LBP texture was reported as more effective than the original LBP.
  • Fingerprint spoofing: Fingerprint methods are broadly hardware-based or software-based and use quality measures, texture descriptors, or learned filters.Examples include ridge strength, directionality, LPQ, WLD, MBLTP, and BSIF across LivDet benchmarks.
  • Fingerprint spoofing: Fingerprint spoofing detection remains an open problem, with LivDet 2013 results still far from perfect.The paper follows a data-driven trend that learns features directly from acquisition data with little prior knowledge.

D. Multi-modalities

The paper positions its method against general image-quality-feature approaches and evaluates it on publicly available iris, face, and fingerprint benchmarks. These datasets include printed, displayed, video, mask, and artificial-replica attacks.

  • Cross-modal comparison: A prior general approach uses 25 image-quality features across face, iris, and fingerprint spoofing, whereas this work learns features with architecture and filter optimization.The paper evaluates on more recent and updated benchmarks than that prior work.
  • Benchmarks: The considered benchmarks are publicly available and suitable for evaluating countermeasure methods against iris, face, and fingerprint spoofing attacks.Table I summarizes their major characteristics.
  • Iris benchmarks: Iris benchmarks include live and printed images captured with the same or specified acquisition sensors, including printer-generated attacks.BioSec, Warsaw, and MobBIOfake use printed-image spoofing constructions.

C. Fingerprint Spoofing Benchmarks

The fingerprint benchmarks include real and fake readings from four sensors, with attacks produced under both cooperative and non-cooperative conditions. The methodology combines architecture search and learned representations to address unclear discriminative features across variable spoofing images.

  • Four sensors contribute fingerprint readings: Biometrika FX2000, Italdata ET10, Crossmatch L Scan Guardian, and Swipe.
  • Fake fingerprints were generated with and without user cooperation, using materials including gelatin, silicone, and latex.
  • Fingerprint images vary substantially across and within modalities, while discriminative real-versus-spoof features remain unclear.
  • Candidate architectures extract training representations and are evaluated with linear SVM scores until the best valid architecture is returned.The procedure repeatedly samples hyperparameters, extracts features, evaluates the objective, and stops after a termination criterion.
  • Randomly generated, normalized filters with rectified activations discard about 50% of expected responses and are reported to improve feature-extraction robustness.

B. Filter Optimization (FO)

Filter optimization learns convolutional filter weights by back-propagation on predefined networks, first using cf10-11 and then a spoofing-specific architecture. The resulting spoofnet modifies cf10-11 to reduce parameters and accept larger inputs.

  • Filter optimization learns network filter weights through back-propagation after an architecture has already been defined.
  • cf10-11 is a three-layer convolutional network whose final stage includes unshared local filtering, a fully connected sublayer, and softmax regression.
  • Spoofnet reduces the first-layer filters from 64 to 16, reorders second-layer operations, and removes third-layer unshared filtering.The authors describe the removed operations as inappropriate where object structure is irrelevant.
  • These changes reduce network weights and allow spoofnet to increase input size from 32 × 32 to 128 × 128.The larger input is intended to improve sensitivity to subtle local image patterns.
  • Both networks use data augmentation during training through cropped image regions and horizontal reflections.

C. Elementary Preprocessing

Preprocessing adapts face and fingerprint images to their acquisition characteristics before representation learning. Face videos are sampled and cropped, while fingerprint images are sensor-specific and remove background or blank-row content.

  • Face videos are evenly subsampled into 10 frames, face positions are detected, and 200 × 200 regions are cropped around them.
  • Fingerprint preprocessing is defined by sensor type because sensors capture different image sizes and background amounts.
  • Swipe images remove variable blank bottom rows using the training-set average number of non-blank rows before resizing and central cropping.
  • Fingerprint crops discard background content because LivDet2013 images often contain large background regions that could mislead representation learning.
  • Inputs are resized to 32×32 for cf10-11 or 128×128 for spoofnet, with color retained according to the AO or FO procedure.
  • Evaluation uses ACC and HTER after obtaining a threshold τ from development predictions, AO cross-validation scores, or τ = 0.5 for FO probabilities.

E. Implementation

The implementation evaluates AO, FO, and spoofnet across nine iris, face, and fingerprint benchmarks, comparing them with state-of-the-art methods. Together, the approaches outperform prior results in eight benchmarks, while performance depends on modality and training-data conditions.

  • E. Implementation: AO implementation probes 2,000 candidate architectures and takes less than one day on average for a benchmark using an Intel i7 and Tesla K40.
  • Results: Across nine benchmarks, AO outperforms four state-of-the-art methods and performs on par in four additional benchmarks, except Swipe.
  • Results: AO architectures often retain original image size: this occurs for all iris and three of four fingerprint benchmarks, while face preferences differ.
  • Limitations: FO performance can be limited by small training sets or irrelevant training structure, and fewer training examples reduce spoofnet generalization.Parameter tuning may reduce the small-sample impact, but the authors identify it as an open research topic.
  • Results: Spoofnet outperforms state of the art in five of nine benchmarks, extending gains where cf10-11 already performed well.The authors associate this result with a simpler architecture, fewer parameters, and a better-suited input size.
  • Results: Combining AO and FO outperforms state-of-the-art methods in eight of nine benchmarks; Biosec is the sole exception.

C. Interplay between AO and FO

The experiments combine architecture optimization (AO), filter optimization (FO), and random filters to examine how these choices interact across biometric spoofing benchmarks. Combined approaches perform especially strongly on fingerprint benchmarks, while random filters can recover performance where learned filters struggle.

  • AO+FO exceeds previous state of the art in all fingerprint benchmarks and performs remarkably well in MobBIOFake.
  • The face and two iris benchmarks retain the difficulty previously observed with FO when AO and FO are combined.
  • AO may outperform spoofnet slightly in cases where the state of the art is exceeded.
  • Random filters improve the overall performance balance of cf10-11 and spoofnet across the benchmarks.
  • spoofnet with random filters exceeds previous state of the art in Replay-Attack.
  • The Replay-Attack weakness of spoofnet with learned filters was not a matter of architecture, because random filters improved performance there.

D. Runtime

The paper estimates processing costs for convolutional spoofing systems and examines samples near real-fake decision boundaries. Runtime depends on the benchmark and implementation, while visual inspection reveals clearer patterns for some face cases than for fingerprint cases.

  • Runtime: 148 (97+51) seconds are required to process 2,200 Swipe images using the most computationally intensive deep representation.
  • Runtime: The runtime estimate relies on a GPU+CPU implementation, which is described as critical for this learning task.
  • Runtime: A hypothetical operational scenario can ignore the 51-second classifier-training cost.
  • Runtime: A single image captured by a Swipe sensor is estimated to require approximately 45 milliseconds plus minor overhead for full processing.
  • Boundary samples: In 3DMAD, real and fake hit samples are visually distinguishable, and no misses were observed among the described hits.
  • Boundary samples: Visual inspection is harder for fingerprint benchmarks because Swipe images are narrow and distorted during acquisition.

VI. CONCLUSIONS AND FUTURE WORK

The study combines architecture optimization and filter-weight learning for biometric spoofing detection, examining their interplay across iris, face, and fingerprint benchmarks. The approaches achieved strong results, while conclusions identify dataset size, modality scope, and future representation strategies as important boundaries and directions.

  • Approaches: The study investigated architecture optimization and filter-weight learning, including their combination in the adapted spoofnet network.The two approaches learn representations through network architecture design and filter weights, respectively.
  • Results: The combined approaches achieved outstanding classification results across the evaluated spoofing problems and outperformed state-of-the-art results in eight of nine benchmarks.The BioSec benchmark was the only case without a state-of-the-art result; 98.93% accuracy was reported against 100.0% in the literature.
  • Results: Learning directly from data produced discriminative fingerprint filters that explored blurring artifacts caused by recapture.The observation was consistent with previous custom-tailored solutions.
  • Approaches: Filter learning was supported by fingerprint benchmarks with a few thousand training samples, whereas small sample size and subject variability hindered it for faces and two iris benchmarks.Architecture optimization remained comparable to state-of-the-art results in almost all benchmarks and outperformed them in three of four cases where filter learning failed.
  • Future work: The proposed approaches may be adapted to other biometric modalities, and future work includes deep representations over pre-processed image feature maps and temporal coherence for video attacks.The authors caution that the reported results are not the final word in spoofing detection.
  • Future work: The iris evaluation covered printed attacks, while cosmetic contact-lens datasets were identified as an available future evaluation target.The authors also propose evaluating other biometric modalities, including palm, vein, and gait.

APPENDIX A

The appendix presents convolutional-network operations as image-processing steps that transform multiband images into progressively richer feature representations. These operations include convolution, rectification, spatial pooling, and divisive normalization.

  • Stacked convolutional operations extract higher-level representations called multiband images, whose pixel attributes form high-dimensional feature vectors for pattern recognition.
  • 1) Filter Bank Convolution: A multiband filter bank applies filters to an input image, producing one output band for each filter.
  • 2) Rectified Linear Units: Rectified linear units replace each filter response with its maximum against zero.
  • 3) Spatial Pooling: Spatial pooling aggregates activations within regions and subsamples them with stride s, providing translational invariance; when s = 2, the output has one-quarter as many pixels.
  • 4) Divisive Normalization: Divisive normalization uses neighboring pooled filter bands to strengthen high responses relative to low ones, increasing output-representation robustness.
Loading 1410.1980v3…