Abstract
Facial recognition and verification is a widely used biometric technology in security systems. Unfortunately, face biometrics is vulnerable to spoofing attacks using photographs or videos. In this paper, we present an anisotropic diffusion-based kernel matrix model (ADKMM) for face liveness detection to prevent face spoofing attacks. We use anisotropic diffusion to enhance the edges and boundary locations of a face image, and the kernel matrix model to extract face image features which we call the diffusion-kernel (D-K) features. The D-K features reflect the inner correlation of the face image sequence. We introduce convolution neural networks to extract deep features, and then employ a generalized multiple kernel learning method to fuse the D-K features and the deep features to achieve better performance. Our experimental evaluation on two publicly available datasets shows that the proposed method outperforms state-of-the-art face liveness detection methods.
Introduction
Biometric security, particularly face recognition, has become ubiquitous in modern technology, ranging from personal device unlocking to high-security access control. However, these systems face a significant threat from "spoofing" attacks, where an unauthorized user attempts to gain access using a 2D representation of an authorized user, such as a printed photograph or a video replay on a screen.
The challenge lies in Face Liveness Detection: the ability of a system to distinguish between a genuine 3D human face and a 2D spoof. While humans can intuitively detect differences in texture and light reflection, automating this process reliably across varying environments remains a complex computer vision problem.
Methodology: The ADKMM Pipeline
The proposed Anisotropic Diffusion-based Kernel Matrix Model (ADKMM) addresses this challenge through a multi-stage feature extraction and fusion pipeline. The architecture combines physics-based image processing with deep learning to capture both structural and high-level semantic information.
The pipeline consists of three primary components:
- Anisotropic Diffusion: This stage acts as a "smart filter." Unlike standard isotropic blurring, anisotropic diffusion preserves and sharpens edges while smoothing flat regions. This process highlights the subtle surface properties and depth cues that differ between real skin and flat spoofing materials.
- Kernel Matrix Model (D-K Features): Following diffusion, "Diffusion-Kernel" (D-K) features are extracted using a Radial Basis Function (RBF) kernel. These features model the non-linear relationships and inner correlations within the image sequence.
- Deep Feature Extraction & Fusion: To complement the structural D-K features, a pre-trained Convolutional Neural Network (CNN) extracts high-level deep features from the diffused images. Finally, Generalized Multiple Kernel Learning (GMKL) is used to fuse these two feature sets, allowing the model to leverage the strengths of both classical and deep learning approaches.
Experimental Evaluation
The model was rigorously tested on two benchmark datasets: the NUAA Dataset (containing 12,614 images of live and photographed faces) and the Replay-Attack Dataset (consisting of 1,300 video clips covering various attack types like print, mobile, and tablet displays).
A key finding in the methodology was the impact of the number of diffusion iterations ($L$). The researchers discovered a "sweet spot" where the diffusion process optimally balances noise reduction and edge preservation.
As shown in the experimental results, the system achieved its peak performance of 99.3% accuracy at $L=15$. Beyond this point, excessive diffusion began to degrade the critical edge information, leading to a decrease in accuracy.
Key Results and Insights
The ADKMM method demonstrated superior performance compared to existing state-of-the-art (SOTA) techniques. On the NUAA dataset, it reached 99.3% accuracy, surpassing methods like ND-CNN (99.0%) and DS-LSP (98.5%). More significantly, on the challenging Replay-Attack dataset, ADKMM achieved a Half Total Error Rate (HTER) of 4.30%, a substantial improvement over the 10% HTER reported for previous SOTA models.
The success of ADKMM stems from its hybrid nature. By using the heat equation (diffusion) to amplify textural differences before feature extraction, the model gains a physics-informed perspective on the scene. The fusion of these structural D-K features with CNN-derived deep features provides a robust defense against diverse attack vectors, from matte paper prints to high-definition digital screens.
Conclusion
The ADKMM framework represents a significant advancement in face liveness detection by effectively combining anisotropic diffusion with kernel-based and deep learning features. While the current results are highly promising, future work could explore the model's computational efficiency for real-time mobile applications and its robustness against sophisticated 3D mask attacks. Nonetheless, the integration of physics-based preprocessing with modern AI provides a powerful template for more secure biometric systems.
Summary updated