DeepScan: A Heuristic-Based Framework for Deepfake and AI-Generated Image Detection Without Neural Network Inference

Abstract: As AI-generated imagery becomes increasingly difficult to distinguish from authentic photographs, the need for accessible detection tools has never been greater. This paper presents DeepScan, a lightweight heuristic-driven image analysis framework that identifies AI-generated or face-swapped images without relying on any pre-trained neural network or GPU hardware. DeepScan applies six calibrated visual heuristics — skin pixel ratio, dark region density, centre-to-background sharpness differential, colour palette diversity, face-region noise estimation, and Error Level Analysis — and combines them through a weighted scoring mechanism to produce a composite authenticity score. The system outputs one of three verdicts: Likely Real, Uncertain, or Likely Fake. Testing across a diverse set of AI-generated portraits and real-world photographs shows a mean fake score of 75.9% for synthetic faces and 8.3% for authentic images, demonstrating strong class separation. Built with Python, Pillow, NumPy, and Flask, DeepScan requires no training phase and consumes minimal computational resources. It is deployed as a REST API accessible through any web browser, making it immediately practical for journalists, media platforms, and content moderation teams.

Keywords: Deepfake Detection, AI-Generated Image Analysis, Error Level Analysis, Heuristic Image Forensics, Skin Pixel Ratio, Colour Diversity, Face Noise Estimation, Image Authenticity, Media Forensics, Flask API, Synthetic Media Detection