ID Card Spoofing Detection Using Frequency Features and CNNs

Abstract

This paper proposes a lightweight and robust approach for detecting spoofed ID card images by integrating convolutional neural networks (CNNs) with frequency-domain analysis. The model adopts a dual-branch design: one branch processes the original RGB image, while the other takes a frequency-enhanced version produced using a high-pass Fast Fourier Transform (FFT) filter. Both branches use the same architecture: the first seven convolutional layers of the VGG16 backbone but each branch has its own parameters. The two streams are merged by a multi-head cross-attention fusion module, which aligns and integrates the complementary cues from both branches more effectively, followed by a classification module for “genuine” vs. “spoof”. The method is evaluated on the "or" and "re" subsets of the Document Liveness Challenge 2021 dataset (DLC-2021). On these subsets, the model attains precision of 93.64%, recall of 88.90% and accuracy of 91.68%, significantly outperforming baseline models. The implementation remains computationally efficient, requiring about 0.120 s per image on an Intel Xeon 2.20 GHz (x86-64) CPU. The approach achieves a favorable trade-off by combining high detection accuracy with a compact model size. These results highlight the benefit of exploiting both spatial and frequency features to enhance the reliability of electronic identity verification systems.