Штучний інтелект

Анотація: Face anti-spoofing (FAS) on mobile devices must generalize across domains under tight constraints on speed, weight, accuracy, and power (SWAP). Recent vision-language approaches such as FLIP achieve strong cross-domain accuracy using ~86 M-parameter CLIP-ViT backbones that are an order of magnitude too large for mobile deployment. We close this gap by distilling a frozen FLIP-MCL teacher into a ~1.6 M-parameter MobileNetV3-Small student using a composite objective combining logit KL divergence, prompt-conditioned feature alignment, and an SSDG asymmetric triplet on student embeddings. Trained under a CIM-style multi-source protocol on OULU-NPU, Replay-Attack and CelebA-Spoof, the student reaches 18.92±0.84% average cross-domain ACER on the OULU held-out split (-24.78 pp over a strong test-time-adaptation baseline at 43.70%) while preserving sub-2 ms latency on flagship Android phones and sub-6 ms on entry-tier devices. An ablation shows the teacher contributes -9.26 to -13.01 pp ACER on top of source-only cross-entropy training and collapses seed variance by 8×, confirming that distillation rather than supervised training drives the gain.

Ключові слова: face anti-spoofing, knowledge distillation, domain generalization, vision-language models, mobile deployment, neural networks, SWAP

Посилання:

1. Srivatsan, K., Naseer, M., & Nandakumar, K. (2023). FLIP: Cross-domain face anti-spoofing with language guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 19685–19696). https://doi.org/10.1109/ICCV51070.2023.01803
2. Liu, A., Xue, S., Gan, J., Wan, J., Liang, Y., Deng, J., Escalera, S., & Lei, Z. (2024). CFPL-FAS: Class free prompt learning for generalizable face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 222–232). https://doi.org/10.1109/CVPR52733.2024.00029
3. 3. Stets, O., & Konovalenko, I. (2026). Resource-aware adaptation at test time for mobile face anti-spoofing under SWAP constraints. Optical-Electronic Information-Energy Technologies, 51(1), 79–89. https://doi.org/10.31649/1681-7893-2026-51-1-79-89
4. Jia, Y., Zhang, J., Shan, S., & Chen, X. (2020). Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 8481–8490). https://doi.org/10.1109/CVPR42600.2020.00851
5. Wang, C.-Y., Lu, Y.-D., Yang, S.-T., & Lai, S.-H. (2022). PatchNet: A simple face anti-spoofing framework via fine-grained patch recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 20281–20290). https://doi.org/10.1109/CVPR52688.2022.01964
6. Le, B. M., & Woo, S. S. (2024). Gradient alignment for cross-domain face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 188–199). https://doi.org/10.48550/arXiv.2402.18817
7. Zhou, Q., Zhang, K.-Y., Yao, T., Lu, X., Ding, S., & Ma, L. (2024). Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 175–187). https://doi.org/10.48550/arXiv.2403.19334
8. Liu, S., Wang, Q., & Yuen, P. C. (2024). Bottom-up domain prompt tuning for generalized face anti-spoofing. In Computer Vision – ECCV 2024 (pp. 170–187). Cham: Springer. https://doi.org/10.1007/978-3-031-72897-6_10
9. Kong, Z., Zhang, W., Wang, T., Zhang, K., Li, Y., Tang, X., & Luo, W. (2024). Dual teacher knowledge distillation with domain alignment for face anti-spoofing. IEEE Transactions on Circuits and Systems for Video Technology, 34(12), 13177–13189. https://doi.org/10.48550/arXiv.2401.01102
10. Stets, O., & Konovalenko, I. (2025). PI-KD: Privileged-information multi-teacher distillation for mobile face anti-spoofing. In Proceedings of AdvAIT 2025. https://ceur-ws.org/Vol-4163/short4.pdf
11. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML) (Vol. 139, pp. 8748–8763). https://doi.org/10.48550/arXiv.2103.00020
12. Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. https://doi.org/10.48550/arXiv.1503.02531
13. Boulkenafet, Z., Komulainen, J., Li, L., Feng, X., & Hadid, A. (2017). OULU-NPU: A mobile face presentation attack database with real-world variations. In Proceedings of the 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG) (pp. 612–618). https://doi.org/10.1109/FG.2017.77
14. Chingovska, I., Anjos, A., & Marcel, S. (2012). On the effectiveness of local binary patterns in face anti-spoofing. In Proceedings of the International Conference of the Biometrics Special Interest Group (BIOSIG) (pp. 1–7). https://publications.idiap.ch/index.php/publications/show/2447
15. Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., & Liu, Z. (2020). CelebA-Spoof: Large-scale face anti-spoofing dataset with rich annotations. In Computer Vision – ECCV 2020 (pp. 70–85). Cham: Springer. https://doi.org/10.48550/arXiv.2007.12342