Our project tackles the "Robust Detection of AI-Generated Images Under Real-World Transformations" challenge: distinguishing authentic photos from AI-generated images. staying accurate after the kind of processing images actually go through online — JPEG re-compression, blurring, resizing, noise, color adjustment, and cropping.
The core insight driving our approach: naive AIGC classifiers tend to key off high-frequency generator artifacts (GAN checkerboarding, diffusion noise residuals) that are exactly what these real-world transforms destroy — which is why clean-only-trained models often collapse once an image has been re-uploaded, cropped, or filtered. Rather than only testing robustness after training, we bake the challenge's exact transform spec (JPEG @ quality 90/70/50/30, Gaussian blur @ σ 0.5/1.0/2.0, 0.5×/0.25× downscale-upscale, Gaussian noise @ σ 0.02/0.05/0.10, ±20% color jitter, 80% center crop) directly into training-time augmentation — each training image is randomly hit with 0–2 of these transforms at a randomly sampled severity, forcing the model to learn features that survive them rather than memorizing artifacts that vanish under compression.
We fine-tune a pretrained EfficientNet-B0 backbone with a custom binary classification head on top, evaluate it under every transform × severity combination from the spec table (plus a clean baseline) to produce a robustness comparison table, and ship a standalone inference script that scores any directory of images and outputs a confidence score per image in the required JSON format.
Log in or sign up for Devpost to join the conversation.