PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images

PRADA addresses a specific gap in the synthetic media detection landscape: images generated by autoregressive (AR) models (e.g., LlamaGen, VAR, Parti) rather than the more commonly studied diffusion models. Its technical approach relies on probability-ratio tests, a technique widely used in NLP for LLM-generated text detection (similar to DetectGPT), now adapted for the visual token space. From a competitive standpoint, the project has low defensibility (score: 3) because it is a reference implementation of a research paper with zero stars and minimal community traction yet. The 'moat' is purely the specific weights or thresholds derived from the research, which are easily reproducible once the paper is public. Frontier labs (OpenAI, Google) pose a high risk because they own the models being detected; they have access to the full log-probability distributions of their internal models, allowing them to implement significantly more robust internal watermarking or detection (like Google's SynthID) that would render third-party black-box tools like PRADA obsolete. Furthermore, the displacement horizon is short (6 months) because detection is a cat-and-mouse game; as soon as a new version of an AR generator is released with a different vocabulary or architecture, this specific implementation likely requires retraining or re-calibration.