TechJam2026Track5

How the solution addresses the problem statement

The detector combines RGB content with an additional frequency-domain channel. An ImageNet-pretrained EfficientNet-B0 is modified to accept four channels and is fine-tuned as a binary classifier. The training pipeline applies label-independent JPEG, crop, color, blur, noise, and resize augmentations to reduce reliance on dataset-specific compression artifacts. A validation-derived threshold is stored in the checkpoint to control the false-positive rate on real images.