Deep Generative And Discriminative Modeling For Image Forgery Detection And Localization

Abstract

Digital images are widely used in domains such as journalism, healthcare, law enforcement, and social media, where authenticity is crucial [1], [2]. However, the rapid evolution of image editing tools and AI-driven generative models has enabled highly realistic manipulations, including copy–move, splicing, and deepfake generation [3], [10]–[12]. These sophisticated forgeries are increasingly difficult to detect, making traditional image forensic methods inadequate for modern scenarios [1], [2], [11]. Despite advancements in deep learning-based forensics, significant challenges remain [6]–[8]. Existing approaches typically rely on either generative models [3]–[5], which reconstruct image distributions but lack strong classification and localization capabilities, or discriminative models [6]–[8], [13], [14], which perform well in classification but struggle with precise localization. Furthermore, current datasets are limited in diversity and often fail to represent real-world manipulations, affecting model generalization [10], [17], [18]. This research proposes a unified deep learning framework that integrates generative and discriminative models for robust image forgery detection and localization. A GAN-based generative module performs reconstruction-driven anomaly detection by learning authentic image patterns and producing residual maps that highlight manipulated regions [3]–[5]. In parallel, a Vision Transformer (ViT)-based discriminative module utilizes selfattention mechanisms to capture global dependencies, enabling accurate classification and pixel-level localization through attention heatmaps [13], [14], [19], [20]. By combining reconstructionbased anomaly detection with attention-driven analysis, the framework enables simultaneous detection and localization while providing explainable outputs [8], [9], [16]. This work contributes to advanced digital image forensics by delivering a scalable, robust, and interpretable solution for detecting AI-generated manipulations, with applications in media verification, cybersecurity, and law enforcement [11], [12], [16], [20].