AI-Based Image Forgery Detection System Using EfficientNet-B0 and Grad-CAM Explanations with a Web-Based Interface

Digital photographs are now routinely used as evidence in news reporting, courts, insurance claims and identity checks, yet inexpensive editing software and generative models make it easy to alter them without visible traces. Manual inspection by experts does not scale, and many classical detectors fail once an image has been recompressed or resized. This paper presents the design and prototype implementation of an AI-based image forgery detection system that labels an uploaded picture as genuine or forged, reports a confidence value, and shows a Grad-CAM heat map of the regions that most influenced the decision. The system couples a Streamlit front end with a Flask back-end service. Uploaded images are resized to 224×224 pixels, smoothed with a Gaussian filter (σ = 0.5), normalised in the manner expected by EfficientNet, and passed to an EfficientNet-B0 classifier. A history page and a downloadable PDF report are provided for registered users. The prototype was exercised on sample inputs, and the captured screens show end-to-end operation, including a low-margin decision (56.43 % confidence) whose Grad-CAM map illustrates why a confidence score alone is not sufficient for forensic use. A quantitative evaluation protocol on public tampering benchmarks is specified as the next phase of the work.