Face Presentation Attack Detection via Content-Adaptive Spatial Operators

A new research paper introduces CASO-PAD, a breakthrough model for detecting facial presentation attacks—a critical security threat where attackers use prints, replays, or masks to spoof facial authentication systems. Developed by researcher Shujaat Khan, the model proposes a novel architectural enhancement to the lightweight MobileNetV3 backbone by integrating content-adaptive spatial operators, a technique known as involution. Unlike standard convolutional kernels that share weights across spatial locations, these operators generate unique, location-specific kernels conditioned on the input content itself. This allows the model to better capture highly localized spoof artifacts, such as subtle texture inconsistencies or lighting anomalies, with minimal computational overhead.

The technical specifications make it suitable for mobile deployment: CASO-PAD contains only 3.6 million parameters and requires just 0.64 GFLOPs to process a 256x256 pixel image. It operates on a single RGB frame, eliminating the need for temporal analysis or auxiliary hardware like depth or infrared sensors. In rigorous testing across major benchmarks, the model demonstrated exceptional performance. It achieved perfect 100% accuracy and a 0.00% Half-Total Error Rate (HTER) on the Replay-Attack and Replay-Mobile datasets. On the more challenging large-scale SiW-Mv2 benchmark (Protocol-1), it attained 95.45% accuracy with a 3.11% HTER, indicating strong robustness against diverse, real-world attack scenarios.

The implications are significant for the security of biometric systems. CASO-PAD provides a practical, efficient pathway to deploy robust anti-spoofing directly on edge devices like smartphones, laptops, and access control terminals. Its single-frame, RGB-only approach reduces hardware costs and complexity while maintaining state-of-the-art detection rates. This advancement directly addresses a growing vulnerability as facial recognition becomes ubiquitous, offering a powerful tool to prevent unauthorized access through sophisticated spoofing attempts.