Precise and Privacy-Preserving Tampering Detection Model for Document Images

Document images, which are crucial vital for information dissemination, often contain sensitive data such as ID and bank card numbers. If this information is tampered with or leaked, it will lead to severe economic losses for individuals and businesses. Moreover, the private nature of these images results in data silos that hinder model learning. Federated learning (FL) provides a good solution by enabling collaborative model training without data sharing, but attackers can still infer private document images from local model gradients using methods like Deep Leakage from Gradients (DLG). Towards this end, we propose a Precise and Privacy-preserving Tampering Detection model for document images under FL framework, named PPTD. In PPTD, a multi-scale and multi-domain image tampering detection model is designed to achieve precise detection and localization of tampered regions. Besides, we construct a fine-grained differential privacy scheme that combines an adaptive privacy budget allocation strategy and a client selection strategy to prevent gradient leakage with high detection accuracy. Finally, the experiments on the T-SROIE and DocTamper datasets demonstrate that the proposed PPTD is effective in resisting DLG attacks, which addresses the ticket fraud issues, and is suitable for deployment in real financial scenarios.