[Paper Review] On the Learning of Deep Local Features for Robust Face Spoofing Detection

Gustavo Botelho de Souza, João Paulo Papa+1 more|arXiv (Cornell University)|Jun 19, 2018

Biometric Identification and Security4 citations

TL;DR

This paper proposes a two-stage CNN architecture, lsCNN, that first pre-trains on local facial regions to learn discriminative spoofing cues before fine-tuning on full faces. By leveraging region-specific deep features, the method achieves state-of-the-art performance on Replay-Attack and CASIA FASD datasets with faster convergence and improved robustness compared to holistic or random patch-based training approaches.

ABSTRACT

Biometrics emerged as a robust solution for security systems. However, given the dissemination of biometric applications, criminals are developing techniques to circumvent them by simulating physical or behavioral traits of legal users (spoofing attacks). Despite face being a promising characteristic due to its universality, acceptability and presence of cameras almost everywhere, face recognition systems are extremely vulnerable to such frauds since they can be easily fooled with common printed facial photographs. State-of-the-art approaches, based on Convolutional Neural Networks (CNNs), present good results in face spoofing detection. However, these methods do not consider the importance of learning deep local features from each facial region, even though it is known from face recognition that each facial region presents different visual aspects, which can also be exploited for face spoofing detection. In this work we propose a novel CNN architecture trained in two steps for such task. Initially, each part of the neural network learns features from a given facial region. Afterwards, the whole model is fine-tuned on the whole facial images. Results show that such pre-training step allows the CNN to learn different local spoofing cues, improving the performance and the convergence speed of the final model, outperforming the state-of-the-art approaches.

Motivation & Objective

• To address the vulnerability of face recognition systems to spoofing attacks using printed photos or digital replicas.
• To investigate whether learning deep local features from distinct facial regions improves spoofing detection performance.
• To develop a training strategy that enhances feature learning efficiency and model generalization in spoofing detection.
• To outperform existing CNN-based methods by exploiting region-specific spoofing cues through structured pre-training.

Proposed method

• The method uses a two-stage training process: first, separate 'PatchNets' are trained on fixed, predefined facial regions (e.g., eyes, nose, mouth) to learn local spoofing cues.
• Each PatchNet is a small, dedicated CNN that learns deep features specific to its assigned facial region.
• After local pre-training, the weights from all PatchNets are combined into a unified 'lsCNN' model for global fine-tuning on full-face images.
• The fine-tuning stage uses standard backpropagation with a base learning rate of 0.0001, momentum parameters of 0.9 and 0.999, and a batch size of 64.
• The architecture is implemented using the Caffe framework, with training on both Replay-Attack and CASIA FASD datasets.
• The final model is evaluated using Equal Error Rate (EER) and Half-Total Error Rate (HTER) on validation and test sets.

Experimental results

Research questions

• RQ1 Can pre-training on local facial regions improve the detection of spoofing attacks compared to holistic or random patch-based training?
• RQ2 Does learning region-specific spoofing cues lead to faster convergence and better generalization in deep CNNs for face spoofing detection?
• RQ3 How does the performance of lsCNN compare to state-of-the-art methods in terms of EER and HTER on standard benchmark datasets?
• RQ4 Can the two-stage training strategy be applied effectively to compact CNN architectures without requiring large, complex models?

Key findings

• On the Replay-Attack dataset, lsCNN achieved an EER of 0.33% and an HTER of 2.50%, outperforming the traditionally trained lsCNN (HTER: 1.75%) and matching the best-performing methods.
• The lsCNN model reached its best performance at iteration 53,600, significantly earlier than the traditionally trained CNN, which required 74,200 iterations to achieve its best result.
• On the CASIA FASD dataset, lsCNN achieved an EER of 4.44%, matching the best-performing methods, including those using large architectures like VGG-Face.
• The lsCNN model converged much faster than the traditionally trained version, reaching peak performance at iteration 9,800 compared to 80,900 for the baseline.
• The local pre-training step improved the model's ability to learn discriminative spoofing cues, leading to better convergence and robustness.
• The method achieved state-of-the-art performance with a compact CNN architecture, demonstrating that local feature learning enhances both accuracy and training efficiency.