Our project is a robust AI-generated image detector designed for real-world conditions, where images are often compressed, blurred, resized, cropped, filtered, or reposted. We developed a compact four-view ConvNeXt-Tiny model using source-balanced training and consistency learning between clean and transformed images.
The system accepts a directory of images and produces a JSON confidence score indicating how likely each image is AI-generated. Our frozen 27.8M-parameter model achieved 96.76% balanced accuracy and 99.50% AUROC on clean validation data. Across 14 transformation cases in the organiser demonstration, it maintained 91.64% mean balanced accuracy and 98.07% mean AUROC.
Alongside the detector, we built a reproducible evaluation pipeline with identity-safe dataset splits, transformation benchmarking, protected-data safeguards, immutable experiment records, and a checksum-verified public model release.
Log in or sign up for Devpost to join the conversation.