ARCHIVES

The widespread sharing of AI‑generated images on social media platforms demands detection methods that remain accurate after JPEG compression, resizing, and repeated forwarding. Existing detectors often degrade significantly under such real‑world conditions because they learn semantic content rather than generation artifacts. This paper presents a compression‑robust AI‑generated image detector that combines three ideas: aggressive augmentation (JPEG, downscaling, blur), a multi‑branch architecture with bit‑plane (RAID) and pixel residual (PiD) streams, and a novel WebDegraded benchmark that simulates social‑media forwarding (resize → JPEG quality 70 → screenshot). Our augmented baseline achieves 99.80% accuracy on WebDegraded – a 1% improvement over the clean‑trained baseline – demonstrating genuine robustness to compression. For interpretability, the system explains decisions via three forensic signals (bit‑plane noise, residual structure, frequency peaks) that directly relate to generation artifacts. A live Streamlit demo with a compression slider allows users to test robustness interactively. Despite strong performance on CIFAKE and WebDegraded, we observe domain shift when evaluating on unseen generators (e.g., Midjourney), which we discuss as future work. The code, model, and demo are publicly available.