In today's digital landscape, images and videos play a pivotal role as compelling evidence across various domains, including legal proceedings, insurance claims, and social media. However, the widespread accessibility of advanced editing tools raises significant concerns regarding the authenticity of digital content, as manipulations often occur without visual traces. Image forensics professionals are tasked with developing innovative technologies to detect such forgeries effectively. Current research on image forgery detection primarily categorizes methods into three classes: feature descriptor-based detection, inconsistency detection via shadow analysis, and double JPEG compression detection. This paper addresses the pressing challenge of image forgery detection, particularly in real-time applications and online platforms, where traditional methods relying on hand-crafted features, size, and contrast exhibit limitations. We propose a novel fusion-based decision approach leveraging lightweight deep learning models, specifically SqueezeNet, MobileNetV2, and ShuffleNet. Our method is implemented in two phases: initially, we utilize pre-trained weights from these models to assess image forgery. Subsequently, we fine-tune the weights to enhance detection accuracy by comparing results obtained from both the pre-trained and fine-tuned models. Experimental results demonstrate that our fusion-based decision approach significantly outperforms state-of-the-art methods, achieving superior accuracy in detecting image forgeries.

Image forgery, deep learning, Fusion based method.