Fingerprint Forgery Detection Algorithm

• Fingerprint forgery detection algorithm is a methodology that distinguishes live fingerprints from spoof imprints through sensor-specific quality measurements and machine learning classifiers.
• It leverages hybrid pipelines combining quality-measure analysis, CNN-based patch classification, and multi-stream feature fusion to handle noise, occlusion, and sensor variability.
• The system achieves high accuracy and real-time performance on commodity hardware by integrating adaptive thresholds and cross-domain feature extraction techniques.

A fingerprint forgery detection algorithm is designed to discriminate between genuine (live) and artificial (fake) fingerprint imprints, including those produced by elaborate spoofing attacks. The primary goal is robust, automated detection of physical (spoof) or synthetic (GAN-generated) forgeries across a range of sensors, attack materials, and noise/distortion scenarios. Modern approaches encompass quality-measure-based pipelines, advanced deep learning architectures, and recent progress in adversarial robustness and cross-modal fusion.

1. Core Algorithmic Strategies

Fingerprint forgery detection methods deploy a sequence of steps that transform the raw fingerprint image into a decision—live or fake—often by leveraging sensor-specific knowledge and multi-domain features.

Quality-Measure Pipelines:

State-of-the-art quality-measure approaches, as exemplified by "Fingerprint Liveness Detection Based on Quality Measures" (Galbally et al., 2022), segment the foreground via Gabor filters, normalize, and partition into fixed-size blocks. Each block yields local descriptors capturing orientation, ridge flow, frequency, and statistical properties. Ten specific quality measures are defined: Orientation Certainty Level (QOCL), Spectral Energy Concentration (QENERGY), Local Orientation Quality (QLOQ), Continuity of Orientation Field (QCOF), mean and standard deviation, and multiple sinusoid-based ridge-valley fit descriptors. Global features are aggregated and a classifier (Linear Discriminant Analysis, LDA) generates the final decision.

Deep Learning Pipelines:

Recent algorithms employ convolutional neural networks (CNNs) to classify either the entire image (Park et al., 2018), overlapping patches (Park et al., 2018, Kiefer et al., 2023), or multi-stream features (Miao et al., 2023). Some frameworks extract global second-order statistics via Gram matrices (Park et al., 2018), or fuse domain-specific and frequency-domain features (ridge and GAN artifact streams) (Miao et al., 2023). Sensor-agnostic and sensor-specific feature selection or normalization is a pervasive design element.

Hybrid and Robust Pipelines:

Adaptive-thresholding and wavelet-based local binary pattern extractors ("Adaptive thresholding pattern for fingerprint forgery detection" (Farzadpour et al., 19 Nov 2025)) enhance robustness to common distortions—pixel/block occlusion, severe noise—via anisotropic diffusion and multi-scale feature fusion, classified with a radial-basis-function SVM.

2. Mathematical Formulation of Quality Measures

Block-based and global features are formalized as follows (all from (Galbally et al., 2022)):

• $$\text{QOCL}_i = \frac{\lambda_{1,i} - \lambda_{2,i}}{\lambda_{1,i} + \lambda_{2,i}}$$, with $\lambda_{1,i} \geq \lambda_{2,i}$ eigenvalues of the local gradient covariance in block $i$
• $$\text{QENERGY} = 1 - \frac{ -\sum_{r=1}^R p_r \log p_r }{ \log R }$$; $p_r$ are DFT band energy fractions
• $$\text{QLOQ} = 1 - \frac{1}{N} \sum_{i=1}^N (loq_i)/(\pi/2)$$; $loq_i$ is the average angular variation with eight neighbors
• $\text{QMEAN} = (1/|F|)\sum_{(x,y) \in F} I(x,y)$; $$\text{QSTD} = \sqrt{ (1/|F|) \sum (I(x,y) - QMEAN)^2 } / 255$$

Sinusoid-based descriptors evaluate ridge/valley fit amplitude ($QA$), variance ($QVAR = \exp(-\sigma_i^2/\tau)$), and clarity via histogram overlap.

Sensor-specific performance is enhanced through feature subset selection for each device, with each feature dimension standardized using sensor-specific statistics prior to classification.

3. Deep Learning Architectures and Feature Fusion

Patch-based CNNs process the entire image as a fully convolutional network, assigning live/fake/background probabilities and aggregating per-patch liveness scores. These architectures deploy SqueezeNet-inspired Fire modules for memory efficiency (e.g., 0.54–2.0 MB parameter count), combined with threshold determination on the global liveness score (Park et al., 2018).

Gram-based CNNs extract Gram matrices at multiple depths post-fire modules. Each Gram matrix encodes channel-wise correlations of features and is stacked (three depths) before further convolutions and global average pooling. The final two-class softmax yields live/fake predictions (Park et al., 2018).

Hybrid Multi-stream Pipelines such as RFDforFin (Miao et al., 2023) compute ridge-based 1D FFT features from thinned ridge-skeletons and combine these with frequency-domain artifact features extracted via a ConvNet from the log-magnitude 2D FFT. The two feature streams are fused by summation and passed through an MLP for final classification.

4. Robustness to Spoof Materials, Distortions, and Generalization

Table: Example Robustness Results in Forgery Detection Systems

Algorithm / Scenario	Distortion	Achieved Accuracy
Adaptive Thresholding	90% pixel miss	0.88
Adaptive Thresholding	70×70 block miss	0.85
Adaptive Thresholding	AWGN, SNR−30dB	0.96

Block-based approaches leveraging anisotropic diffusion (Perona–Malik) and multi-threshold ATPs (Farzadpour et al., 19 Nov 2025) report 5–8% absolute accuracy gains under extreme distortion, comparative to prior local pattern (LPQ) baselines.

Comprehensive anti-spoofing evaluation in GAN-generated fakes (e.g., SpectralGAN++ and SDN++) is achieved by specifically fusing ridge and artifact domains in RFDforFin, yielding >96%–99% accuracy under adversarial postprocessing (Miao et al., 2023).

Generalization beyond known spoof materials is addressed by sensor/feature adaptation (Galbally et al., 2022) and explicit sensor-agnostic design in local quality feature models (Sharma et al., 2018).

5. Experimental Methodologies and Performance Metrics

Fingerprint forgery detection systems are experimentally validated using public datasets (LivDet 2009, 2011, 2013, 2015), with protocols emphasizing intra-sensor and cross-material evaluation.

Standard metrics include:

• False Acceptance Rate (FAR): proportion of fakes misclassified as live
• False Rejection Rate (FRR): proportion of live images misclassified as fake
• Average Classification Error (ACE): mean of FAR and FRR
• Overall Accuracy: $1-\text{ACE}$

Exemplar performance:

Sensor	FAR (%)	FRR (%)	ACE (%)	Accuracy (%)
Biometrika	2.12	1.54	1.83	98.17
CrossMatch	10.3	11.94	11.12	88.88
Identix	6.40	7.07	6.73	93.27

Patch-based CNNs with optimal thresholding reach ACE ≈ 1.35% (Park et al., 2018), while Gram-based CNNs report ACE ≈ 2.61% (Park et al., 2018). Adaptive ATP-SVM approaches are superior under high occlusion and noise (Farzadpour et al., 19 Nov 2025). For GAN imagery, 100% detection is achieved on clean data; ≥96% under anti-forensic perturbation (Miao et al., 2023).

6. Practical Complexity, Real-Time Viability, and Implementation

Algorithmic pipelines are optimized for real-time operation, with quality-measure-based methods achieving 40–60 ms per image on a 2 GHz CPU, and compact CNN-based detectors processing within 21–124 ms on modern GPUs (Galbally et al., 2022, Park et al., 2018, Park et al., 2018).

All presented systems operate on a single static image, do not require user cooperation beyond one scan, and avoid additional hardware—enabling low-latency integration into commodity biometric authentication platforms.

Computational complexity is typically $O(P)$ for segmentation, $O(N)$ for blockwise feature extraction, $O(d^2)$ for LDA classification, with CNN models mostly bounded by convolutional layer depth and parameter count (Galbally et al., 2022, Park et al., 2018).

7. Limitations and Future Directions

Limitations center around sensitivity to extremely degraded fingerprints, untested cross-sensor domains, and evolving attack paradigms including unseen GAN architectures or occlusion strategies. Performance degradation is observable in scenarios with extreme quality variation or distributional shifts between training and test data (Galbally et al., 2022). A plausible implication is that robust generalization will require fusion with hardware cues, self-supervised or meta-learning frameworks, or additional domain-adaptive modeling.

Further advancement may entail:

• Hybrid fusion of dynamic/time-series and static cues, especially for presentation attack detection under time-varying liveness (Plesh et al., 2021)
• Utilization of one-class anomaly detection from pristine image statistics to flag unforeseen forgery modalities (Mareen et al., 2022)
• Deployment of adaptive, sensor-agnostic feature extractors resilient to future synthetic and physical spoof materials

The field continues to move toward deeper integration of texture, frequency, and physiological signals, with increasing emphasis on adversarial robustness and generalization to unseen forgery types.