Learnable Multi-Level Frequency Decomposition and Hierarchical Attention Mechanism for Generalized Face Presentation Attack Detection

Meiling Fang, Naser Damer, Florian Kirchbuchner, Arjan KuijperWACV 2022

TL;DR

This paper introduces LMFD-PAD, a novel dual-stream deep learning framework for face presentation attack detection (PAD). By combining learnable frequency decomposition with a hierarchical attention mechanism, LMFD-PAD significantly improves generalization to unseen scenarios and datasets, addressing a critical limitation of current PAD systems.

What Problem Does This Paper Address?

Face recognition systems are increasingly ubiquitous, but their security is threatened by "presentation attacks" (PAs) – attempts to fool the system with fake faces (e.g., printed photos, replayed videos, 3D masks). While existing face PAD methods achieve good performance on the datasets they are trained on (intra-dataset), their performance often plummets when encountering new attack types, sensors, or lighting conditions (cross-dataset generalization). This lack of robustness is a major hurdle for deploying secure and reliable face recognition systems in real-world applications. The paper aims to bridge this gap by developing a PAD method that is inherently more generalizable.

Methods

The core of LMFD-PAD is a dual-stream architecture:

  • Multi-Level Frequency Decomposition (MFD): One stream processes the input face image in the frequency domain. Instead of using fixed filters, this approach employs learnable frequency filters to decompose the image into multiple frequency bands (low, middle, high, and a full spectrum). The intuition is that frequency domain features are less sensitive to variations in sensors and illumination, and learnable filters allow the model to adaptively focus on the most discriminative frequency components for attack detection. The Discrete Cosine Transform (DCT) is used for this decomposition.
  • Hierarchical Attention Mechanism (HAM): The second stream processes the original RGB image in the spatial domain. The key innovation here is how features from both streams are integrated. Instead of a simple late fusion, HAM integrates features hierarchically. It uses spatial attention in earlier layers to capture fine-grained texture details and channel attention in later layers to capture more abstract semantic information. This approach leverages the different natures of features at various depths of the CNN.
  • Loss Function: The model uses a combined loss function: Smooth L1 loss for pixel-wise supervision (encouraging accurate feature maps) and Focal Loss for binary supervision (focusing on hard-to-classify samples). This is presented as an improvement over standard Binary Cross-Entropy (BCE) loss for better generalization.

Datasets

The paper evaluates LMFD-PAD on four publicly available face PAD datasets:

  • Oulu-NPU [3]
  • CASIA-MFSD [37]
  • Idiap Replay-Attack [6]
  • MSU-MFSD [29]

These datasets are used in both intra-dataset (training and testing on subsets of the same dataset) and cross-dataset settings (training on multiple datasets and testing on a held-out one). The authors follow established protocols for evaluation, particularly those from the IJCB-MFAS competition.

Key Insights and Ideas

  • Frequency Domain for Robustness: Decomposing images into frequency components offers a more robust representation against variations in capture conditions (sensors, lighting) compared to solely relying on spatial RGB information.
  • Learnable Filters are Key: Using learnable frequency filters, rather than fixed ones, allows the model to adaptively discover the most discriminative frequency bands for PAD, leading to better generalization.
  • Hierarchical Fusion Matters: Integrating features from the spatial and frequency streams at different network depths, guided by attention mechanisms tailored to feature abstraction levels (spatial for texture, channel for semantics), is more effective than simple late fusion.
  • Generalization Beyond Domain Adaptation: The proposed method achieves strong cross-dataset performance even without explicit domain adaptation modules, suggesting its learned features are inherently more generalizable.

Key Results

  • Intra-dataset Performance: LMFD-PAD achieves competitive results on intra-dataset evaluations, comparable to state-of-the-art methods.
  • Cross-dataset Performance: The method demonstrates significant improvements in cross-dataset generalization. For instance, in one setting, it achieved an HTER of 10.48%, outperforming the next best method at 13.89%. It consistently outperforms methods that do not use auxiliary information like depth or rPPG.
  • Ablation Study: An ablation study confirms the benefits of each proposed component: MFD, HAM, and the combined loss function all contribute positively to the cross-dataset performance.

Limitations and Future Work

  • Intra-dataset vs. Depth/rPPG: While LMFD-PAD performs well compared to pixel-wise and binary supervised methods, it scores below depth/rPPG supervised networks in some intra-dataset scenarios. The authors suggest this could be an avenue for future work, potentially by incorporating depth or rPPG information.
  • Computational Cost: The paper mentions the expensive computation cost of NAS-based methods as a consideration, implying that while their method is effective, computational efficiency might be a factor for deployment.
  • Further Exploration of Components: The paper acknowledges that it's "still unclear which part of our model benefits the improved results" and dedicates an ablation study to this, but further in-depth analysis of specific component contributions could be explored.

Summary updated

BibTeX
@inproceedings{fang2022learnable,
  author = {Meiling Fang and Naser Damer and Florian Kirchbuchner and Arjan Kuijper},
  title = {{Learnable Multi-Level Frequency Decomposition and Hierarchical Attention Mechanism for Generalized Face Presentation Attack Detection}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
}
Figure