Novel Architecture for Fake Identity Document Detection

This project presents a new approach for protecting identity documents by combining Deep Q-Learning with a Dual-Stream CNN. Through the literature review, it was found that standard detection models struggle with the unique textures and scanning artifacts found in PAN cards. The proposed solution addresses this by using a Q-Learning agent to actively find and test the model’s weaknesses through smart data augmentation. By focusing on both pixel-level textures and hidden frequency patterns, the system is designed to catch forgeries that often bypass traditional tools. While the full system is still being built, the proposed architecture offers a clear path toward more reliable identity verification. Future development will focus on building the frequency- analysis branch and testing how well the model handles real-world, low-quality document scans.

[1]     M. S. Rana et al., “A Comprehensive Review of Deepfake Detection Techniques: From Traditional ML to Advanced Deep Learning,” IEEE Access, vol. 14, pp. 2450–2472, Jan. 2026. [2]     Y. Zhou et al., “Improving Deepfake Detection with RL-Based Adaptive Data Augmentation,” in Proc. AAAI Conf. on AI, 2026, pp. 4122–4130. [3]     X. Liu and H. Zhang, “Feature Refinement and Enhancement Network (FRENet) for Deepfake Detection,” Image and Vision Computing, vol. 154, 2025. [4]     E. Liang and K. Gupta, “Digital Image Forgery Detection via Adaptive Q-Learning and Error Level Anal- ysis,” J. Digital Forensics, vol. 21, 2026. [5]     A. Singh and R. Verma, “Automated Identity Document Forensics: A Hybrid CNN-OCR Framework,” MDPI Symmetry, vol. 17, 2025. [6]     J. Wang et al., “Localization of Image Forgery in Structured Identity Documents,” IEEE Trans. Inf. Forensics Security, vol. 20, 2023. [7]     Sumsub Identity Fraud Institute, “2025 Identity Fraud Report,” Sumsub Research, White Paper, 2025. [8]     MeitY, “IT Rules 2026,” Government of India, Tech. Rep., Feb. 2026. [9]     J. Umpierrez et al., “Scanned Documents Forensics: Detecting Inserted Characters,” in Proc. IEEE ICIP, 2025. [10] A. Kumar, “HybridNet: Advancing Deepfake Detection,” IEEE Access, vol. 13, 2025.