Abstract
Automatic methods for detecting presentation attacks are essential to ensure the reliable use of facial recognition technology. Most of the methods available in the literature for presentation attack detection (PAD) fails in generalizing to unseen attacks. In recent years, multi-channel methods have been proposed to improve the robustness of PAD systems. Often, only a limited amount of data is available for additional channels, which limits the effectiveness of these methods. In this work, we present a new framework for PAD that uses RGB and depth channels together with a novel loss function. The new architecture uses complementary information from the two modalities while reducing the impact of overfitting. Essentially, a cross-modal focal loss function is proposed to modulate the loss contribution of each channel as a function of the confidence of individual channels. Extensive evaluations in two publicly available datasets demonstrate the effectiveness of the proposed approach.
Introduction
Face recognition technology has become ubiquitous, yet it remains vulnerable to presentation attacks (PAs). These attacks range from simple 2D media, such as printed photos and digital replays, to highly sophisticated 3D silicone masks and makeup-based obfuscation. While standard RGB-based systems are easily deceived, incorporating additional modalities like depth can significantly enhance security. However, multi-channel systems often struggle with two primary challenges: overfitting to specific training patterns and failing to generalize to "unseen" attacks that were not present during the model's training phase.
The goal of this research is to develop a robust framework that effectively leverages complementary information from multiple sensors while preventing the model from becoming "lazy"—a phenomenon where the network relies too heavily on the biases of a single modality.
Methodology
The proposed framework utilizes a two-stream architecture based on the DenseNet-161 backbone. One stream processes standard RGB images, while the other processes corresponding depth maps. Features from both streams are extracted and eventually fused for a joint classification.
Cross-Modal Focal Loss (CMFL)
The core innovation is the Cross-Modal Focal Loss (CMFL) function. In traditional multi-channel training, a model might ignore one sensor if the other provides a simpler path to minimizing error. CMFL addresses this by dynamically modulating the loss contribution of each channel based on its confidence.
Key aspects of the CMFL include:
- Dynamic Modulation: If the RGB channel is highly confident that a sample is an attack, the loss contribution from the depth channel is reduced, preventing the model from forcing a classification when information may be lacking in that specific modality.
- Collaborative Learning: This approach acts as a form of architectural regularization, ensuring that individual branches learn robust, independent representations before late fusion.
- Harmonic Mean Weighting: The loss modulation uses the harmonic mean of probabilities across branches, ensuring the model is not misled by a single high-confidence error.
Datasets and Experimental Setup
The framework was evaluated on two large-scale multi-channel datasets:
- WMCA (Wide Multi-Channel presentation Attack): Includes 2D and 3D attacks from 72 subjects.
- HQ-WMCA (High-Quality WMCA): A more challenging dataset featuring sophisticated attacks like 3D masks, makeup, and tattoos.
To test generalization, the researchers employed a Leave-One-Out (LOO) protocol. In this setup, the model is evaluated on an attack type (e.g., 3D masks) that was entirely excluded from the training set.
Results and Discussion
The CMFL-based approach demonstrated superior performance in detecting unseen attacks compared to state-of-the-art baselines:
- WMCA Dataset: The proposed method achieved an Average Classification Error Rate (ACER) of 7.6%, a significant improvement over the 11.4% achieved by a standard multi-head model using Binary Cross Entropy.
- HQ-WMCA Dataset: The framework reached an ACER of 11.6%, outperforming the popular MC-PixBiS baseline, which recorded a 19.6% error rate.
- Robustness to Missing Modalities: Interestingly, models trained with CMFL performed better even when only a single channel (RGB) was available at test time, indicating that the loss function encourages the learning of more generalizable features.
Conclusion
This work introduces a novel Cross-Modal Focal Loss that effectively addresses the challenges of multi-channel face anti-spoofing. By dynamically weighting sensor contributions, the framework improves generalization to unseen attacks and reduces overfitting. While the system shows promise, cross-database performance remains a challenge due to sensor mismatch between different depth cameras. Future research will focus on improving data augmentation and exploring more complex fusion strategies to further bridge this gap.
Summary updated