Enhancing Identity Documents Security: Developing a Verification Tool to Prevent Spoofing, Tampering, and Forgery

Abstract

Remote onboarding is now vulnerable to AI-created deepfakes and identity fraud, particularly given the increased compliance standards introduced by eIDAS 2.0, which mandates highly robust identity document verification as a crucial component of secure online operations. This paper proposes a multimodal deep learning model that consists of three unified modules designed to overcome these problems. The initial module is a visual similarity analysis to ensure identity with the help of a Siamese Convolutional Neural Network. The second module applies specific fraud detection in different countries with dynamic thresholds to consider the differences in documents and attacks between jurisdictions. The third module separates original and scanned documents with the help of an EfficientNet-B7-based classifier. The system was trained and tested using the MIDV-2020 dataset that consists of 10 country documents (320 real and 320 forged documents per country). Experimental measures show a True Acceptance rate of 99-100%, a False Acceptance rate of 6.0%, and a pipeline latency of 278 ms on CPU and 42 ms on GPU, which can be used to run GUI operation in real time. The proposed multimodal verification framework integrates structural verification, adaptive fraud detection, and presentation attack detection into a single pipeline, ensuring more robust and reliable digital identity verification while meeting regulatory requirements. The next step in work will be INT8 quantization to make it possible to use it on mobile and IoT platforms efficiently.