Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{199060,
        author = {Pandiripalli Sai Ganesh and Palleti Sasidhar and Kunchakuri Dheeraj and Kusuri Karthik},
        title = {A Robust Deep Learning Framework for Face Anti-Spoofing Detection Using Compact Convolutional Neural Networks},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {12250-12266},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=199060},
        abstract = {Face recognition systems have become ubiquitous in our daily lives—from unlocking smartphones with a glance to passing through automated border control gates at international airports. However, this convenience comes with a dangerous vulnerability: presentation attacks, commonly known as spoofing. Attackers can bypass these systems using something as simple as a printed photograph or as sophisticated as a hyper-realistic 3D silicone mask or a deepfake video generated by artificial intelligence. This paper presents a comprehensive study and a practical solution for face anti-spoofing using a compact Convolutional Neural Network (CNN) designed specifically for resource-constrained environments like smartphones and edge devices. We trained and rigorously evaluated our model on two benchmark datasets: the CASIA Face Anti-Spoofing Database (CASIA-FASD) and the Replay-Attack Database. On the CASIA dataset, our model achieves a training accuracy of 93.37% and a validation accuracy of 91.20%. More importantly, it demonstrates a Bona Fide Presentation Classification Error Rate (BPCER) of just 1.62%, meaning that genuine users are almost never incorrectly rejected. However, our Attack Presentation Classification Error Rate (APCER) of 27.54% reveals that sophisticated attacks—particularly high-quality video replays and 3D masks—remain challenging. Our model contains only 405,377 parameters, making it lightweight enough to run in real-time on devices with limited computational resources. Beyond the technical results, this paper offers an honest, human-centered discussion of the real-world challenges that are often overlooked in academic literature: why a liveness system trained on English lip movements fails for a Tamil or Mandarin speaker, why iris recognition is impractical on most smartphones due to resolution limits, and why asking a person with facial paralysis to” smile for the camera” is not just annoying but exclusionary. We also evaluate our model on the Replay-Attack dataset to test cross-database generalization, revealing significant performance degras-dation (accuracy dropping from 91.20% to 67.5%), highlighting the critical challenge of domain shift in real-world deployment. We conclude with a forward-looking roadmap that includes multimodal fusion (combining visible light with thermal and depth sensors), few-shot learning for adapting to new attack types with minimal examples, and privacy-preserving architectures that never transmit raw face images off the user’s device.},
        keywords = {},
        month = {April},
        }