Work · Independent
ACFS
A detector for AI-generated images and text that never gives a score without the evidence behind it — and hashes its report.
What it does
Two independent approaches, because a single model signature is fragile. Fine-tuned ViT/DeiT and RoBERTa/DeBERTa classifiers look for generative artifacts; Error Level Analysis, EXIF anomalies and stylometry look at physical and statistical signals that don't depend on having seen a particular generator.
Decisions
- Explanations are output, not decoration. Grad-CAM heatmaps for images, SHAP per sentence for text.
- Built for evidence. The PDF report carries a SHA-256 of the original artifact. Datasets are versioned with DVC so retraining is reproducible.