Abstract
This paper addresses the challenge of developing robust Face Anti-Spoofing (FAS) models for face recognition systems. Traditional FAS protocols are limited by a lack of diversity in subject identities and environmental conditions, restricting generalization to real-world scenarios. Recent advancements in spoof image synthesis have mitigated data scarcity but still fail to capture the full range of facial attributes and environmental variability needed for effective domain generalization. To address this, we propose a novel framework capable of generating diverse, realistic facial images with text-guided control. We fine-tune Stable Diffusion to extract real facial features and specifically train LoRA layers to capture detailed spoof patterns. Additionally, the text-guided control of attributes helps overcome the lack of diversity seen in previous methods. Extensive experiments demonstrate that our text-to-image-based synthetic data generation significantly enhances the robustness of FAS models, establishing a new benchmark for domain-independent and reliable anti-spoofing systems.
Introduction
Face recognition systems are increasingly ubiquitous, yet they remain vulnerable to presentation attacks—spoofing attempts using printed photos, digital screens, or 3D masks. Effective Face Anti-Spoofing (FAS) models are required to secure these systems. However, existing FAS datasets are often limited to small subject groups (fewer than 50) and controlled laboratory environments. This lack of diversity in race, age, lighting, and background leads to poor domain generalization; models trained in a lab often fail when deployed in unpredictable real-world settings. Furthermore, because spoofing images do not occur naturally, they cannot be easily web-crawled, resulting in chronic data scarcity.
Methodology: The SpoofFusion Framework
To bridge the gap between lab-trained models and real-world application, we propose SpoofFusion, a three-phase pipeline designed to generate high-fidelity synthetic datasets:
- Refining Facial Features: We first fine-tune a Stable Diffusion U-Net on bonafide (real) facial images to ensure the model produces authentic facial textures.
- Specializing in Spoofs via LoRA: Rather than retraining the entire model, we utilize Low-Rank Adaptation (LoRA) layers trained exclusively on spoofing artifacts (e.g., moiré patterns or paper edges). This allows the model to decouple facial identity from spoofing characteristics.
- Text-Guided Synthesis: By combining the refined U-Net with the spoof-specific LoRA, we use diverse text prompts (e.g., "a woman wearing glasses") to generate an almost infinite variety of synthetic images that encompass various demographics and environments.
Experimental Results
We evaluated SpoofFusion using four gold-standard FAS datasets: Oulu-NPU (O), CASIA-MFSD (C), Idiap Replay-attack (I), and MSU-MFSD (M). We also conducted tests on "wild" datasets like CelebA-Spoof and SiW-Mv2 to verify robustness.
Key Performance Metrics
- State-of-the-Art Accuracy: In standard "leave-one-out" benchmarks, SpoofFusion achieved an average Half Total Error Rate (HTER) of 1.36%, a significant improvement over the previous state-of-the-art of 3.01%.
- Cross-Domain Superiority: In a 12-way cross-domain evaluation, the framework reached an average HTER of 2.56%, nearly halving the error rate of the previous best model (4.84%).
- Diversity Over Realism: A striking finding was that models trained solely on synthetic data often outperformed those trained on limited real datasets, suggesting that for FAS, data diversity is more critical than the source's "realness."
Conclusion and Future Work
SpoofFusion demonstrates that text-to-image synthesis can effectively overcome the data scarcity and diversity limitations inherent in traditional FAS research. By decoupling identity from spoofing artifacts, we can generate robust models that generalize well to unseen domains. Future work will focus on expanding the framework to include more complex attack vectors, such as 3D silicone masks and deepfakes, and refining prompt engineering to simulate specific sensor distortions. While synthetic data mitigates privacy concerns, the ethical implications of generating realistic human faces remain a priority for ongoing scrutiny.
Summary updated