Lightweight spatial attention pyramid network-based image forgery detection optimized for real-time edge TPU deployment

Abstract

The widespread accessibility of image editing software has made image forgery a considerable threat in journalism, legal contexts, and social media, requiring effective and precise detection techniques. The Authors propose a Spatial Attention Pyramid Network (SAPN) that integrates multi-scale residual feature extraction with an adaptive spatial attention mechanism to tackle the difficulties of identifying subtle and localized alterations. SAPN attains enhanced forgery detection performance and computational efficiency by utilizing hierarchical feature learning and selectively augmenting regions susceptible to manipulation. Extensive experiments conducted on four benchmark datasets illustrate the effectiveness and generalizability of SAPN. On the CASIA V1 dataset, SAPN attains an accuracy of 94% and an AUC of 0.99, outperforming 29 state-of-the-art models. An ablation study further supports the contributions of the pyramid feature extraction and spatial attention modules to the overall performance improvements. Moreover, a lightweight model architecture, containing merely 0.57 million parameters, enables efficient real-time deployment on Edge TPU devices, with an average inference latency of 1.17 s per image. These results proclaim SAPN as a scalable and robust framework for image forgery detection and localization in real-world applications.