AIDE: Hybrid Feature Detector for AI‑Generated Image Detection and the Chameleon Benchmark

In the era of rapid AI content generation, distinguishing AI‑generated images from real photographs has become a critical challenge for forensics, misinformation mitigation, and copyright protection.

To address the lack of high‑quality, realistic benchmarks, the Xiaohongshu algorithm team together with USTC and Shanghai Jiao‑Tong University introduced the fully manually annotated Chameleon benchmark and the AIDE detection method, presented at ICLR 2025.

The Chameleon dataset contains ~26,000 test images (over 150 K AI‑generated and 20 K real images) covering humans, animals, objects and scenes, with resolutions from 720 p to 4 K. All generated images passed a human “Turing test”, ensuring high realism. The dataset was built through multi‑channel collection, rigorous filtering (resolution, unsafe content, duplication, text‑image consistency via CLIP) and double‑blind human annotation, achieving high quality and diversity.

AIDE (AI‑generated Image DEtector with Hybrid Features) combines low‑level pixel statistics and high‑level semantic cues. Its Patchwise Feature Extraction (PFE) module selects high‑ and low‑frequency patches via DCT scoring, processes them with SRM filters and ResNet‑50 backbones. The Semantic Feature Embedding (SFE) module extracts global visual embeddings using a pretrained OpenCLIP model. Features from both modules are concatenated and fed to an MLP classifier.

Experiments on the existing AIGCDetectBenchmark and GenImage datasets, as well as the new Chameleon benchmark, show that AIDE outperforms nine state‑of‑the‑art detectors, improving accuracy by 3.5 % and 4.6 % on the former two datasets. However, performance gaps remain on the more challenging Chameleon benchmark, highlighting the need for further research.

The authors plan to enlarge Chameleon, incorporate more categories and generation models, and further refine the AIDE architecture. A lightweight companion work on synthetic image detection (KDD 2025) is also mentioned.