RISA Recaptured-Image Detection
This repository contains the public deployment weights for a binary classifier that distinguishes single-captured images from images recaptured through a display-camera chain.
The release is a deterministic 17-readout RISA model fitted on all 86 development images from 12 content groups in the Imperial College London DS-05 dataset. It uses a frozen 384-dimensional DeiT-S representation together with 479-dimensional edge-profile and 2,035-dimensional native-texture descriptors. The readouts use a linear kernel and aggregate their two-class logits by coordinate-wise median.
Files
model.safetensors: non-executable model tensors, including full-data standardization statistics, standardized support features, 17 dual coefficient matrices, and readout multipliers.config.json: tensor shapes, class mapping, selected deployment configuration, evaluation boundary, and SHA-256 provenance.LICENSE.md: release terms for the weights.
This is a weights-only release. The raw-image feature-extraction implementation and training code are not included. The model therefore expects feature vectors produced by a compatible RISA feature pipeline; it is not a drop-in Transformers checkpoint.
Evaluation
The defensible performance estimate comes from a frozen four-outer/three-inner StratifiedGroupKFold evaluation in which each content identifier stays entirely within one fold.
| Metric | Grouped out-of-fold result |
|---|---|
| Accuracy | 94.19% |
| Balanced accuracy | 93.57% |
| Macro-F1 | 93.82% |
| ROC-AUC | 96.57% |
The confusion matrix is [[30, 3], [2, 51]], with true classes as rows and predicted classes as columns.
The downloadable weights are a later deployment refit on all 86 development images. The 97.67% score stored in config.json is the grouped configuration-selection score used to choose that full-data deployment configuration. It is not an independent test estimate and should not replace the 94.19% grouped out-of-fold accuracy above.
Intended use
The model is intended for research and technical evaluation of recaptured-image screening. A positive score indicates that an image resembles the recaptured class represented in DS-05; it is not proof of deception, manipulation, or malicious intent.
Limitations
- The training set contains only 86 images and 12 scene-content groups.
- Content grouping controls scene leakage but does not establish cross-device, cross-display, or cross-session generalization.
- The model has not been validated as a standalone forensic decision system.
- Benign display photography, accessibility workflows, and document capture can create the same acquisition route.
- Human review and original evidence should be retained for consequential decisions.
Dataset provenance
The source dataset is the Imperial College London DS-05 recaptured-image dataset. No original dataset images are redistributed in this repository.
Integrity
model.safetensors SHA-256:
CE78A552A7CBF9A6E5F1FF540F3AB0AD4E86CD03A792643C76EEA49753D2C6C5
The serialization round-trip was verified with a maximum absolute logit error of 0.0.
- Downloads last month
- 15