Small CNNs achieve >98% accuracy in black-box image generator attribution
kwangmoo_yi · x · 2026-08-20
The paper 'Scalable Black-Box Model Attribution for Images' introduces RPA, a method that uses a lightweight CNN to reliably identify which generative model produced an image in a strict black-box setting.
Core Methodology:
- Patch Division: Splits images into overlapping 256×256 patches, ensuring resolution invariance.
- Per-Patch Classification: A compact 6M-parameter CNN classifies each patch from raw pixels against candidate generators.
- Aggregation: Combines per-patch softmax scores into an image-level label via overlap-corrected weighted averaging.
Performance:
- Achieves 98.0% accuracy on the 25-class DRAGON benchmark and 92.9% on the 27-class OpenFake benchmark.
- Remains robust to compression, blur, and resizing.
- The learned representation enables unsupervised model lineage recovery, detection of unseen generators, and few-shot adaptation for new models without retraining.
Related event: Small CNN Identifies Image Generators with Over 98% Accuracy(2 posts)→
More from Research
- Weak adversarial networks solve NS-Brinkman equations, outperforming PINNs — FrnkNlsn · 2026-08-20
- Stanford releases quickstart guide for ENCODE GRAMMAR genomics AI tool — anshulkundaje · 2026-08-20
- Peking U & MSRA: BCP optimizes robot replanning timing via lightweight policy — 机器之心 · 2026-08-20
- AI agent stacks blocks in browser physics sim as open embodied-AI arena debuts — NovaCoding · 2026-08-20
- Agent Arena: Million-task leaderboard for real-world AI agents — arena · 2026-08-20
- Robot Achieves One-Shot Imitation with Zero Fine-Tuning — Scobleizer · 2026-08-20