Judicial Evidence Verification by Vision Transformer

Photographs, CCTV footage and phone recordings now sit at the centre of many investigations, yet cheap editing tools and generative AI have made it much easier to alter them convincingly. Checking such material by eye is slow, depends heavily on the examiner, and can miss subtle changes, especially in long recordings. This paper studies how a Vision Transformer (ViT) can be used as one part of a larger forensic workflow for screening video evidence. The framework we describe reads a video with OpenCV, samples frames and attaches a timestamp to each, locates and prepares facial regions, classifies each frame as genuine or possibly manipulated, applies a confidence threshold, produces heat-map explanations where possible, and compiles everything into a structured report for a human examiner.