Year: 2026
Type: dissertation
Abstract: Modern vehicles integrate biometric authentication for driver monitoring and access control, making robust liveness detection a safety-critical requirement. Face anti-spoofing (FAS) protects these systems against presentation attacks that can deceive facial recognition into granting unauthorized access. This thesis addresses in- vehicle FAS through three interconnected works. First, a comprehensive survey reviews deep-learning-based FAS methods for automotive deployment, identifying lightweight architectures and multi-modal sensing as the most promising directions. Second, an empirical benchmark compares six lightweight deep learning models on the VFPAD in-vehicle dataset, establishing EfficientNetV2-B0 as the strongest accuracy–efficiency tradeoff. Third, EGA-FAS is proposed: a novel multi-modal framework combining a stage-selective Ghost-Efficient backbone with an Adaptive Modality Weighting Module for learned, per-sample red-green-blue (RGB), depth, and infrared (IR) fusion. EGA- FAS achieves 1.76% average classification error rate (ACER) on WMCA with only 0.144M parameters (9.46 MFLOPs) and 230.9 FPS on NVIDIA Jetson Orin Nano, outperforming methods 2–3 orders of magnitude larger.
Author: Ruiz, Michael
Institution: Western University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208034747
