Digital Forensic Detection of Forged Driver's Licenses Using Optimized ResNet50 and Handcrafted Feature Baselines

Abstract

Automated forged driver's license detection is important for digital identity verification, but many prior studies are difficult to compare because they use different datasets, preprocessing pipelines, and evaluation protocols. This paper presents a controlled comparison of handcrafted and deep learning pipelines for binary forged driver's license detection using the driver-license subset of the IDNet dataset. Five models are evaluated under identical data splits and preprocessing: HOG-SVM, MobileNetV2, EfficientNetB0, ResNet50, and an optimized ResNet50. The optimized ResNet50 achieved the best overall test performance, with 94% accuracy, 96% forged-class recall, and a ROC-AUC of 0.9777. HOG-SVM achieved lower overall accuracy but offered the fastest inference time, indicating its relevance for lightweight forensic screening. Fraud-type and visualization analyses show that structurally explicit manipulations are easier to detect, while face morphing remains the most challenging category because it preserves global facial continuity. The results suggest that optimized residual learning improves detection robustness, but real-world validation remains necessary because the experiments rely on synthetic identity documents.