datapoisoning-detector-v1-medium

Architecture at a glance

Original diagrams for this release. Editable SVG versions: architecture · inference and distillation.

Weights and completed test results are included in this release bundle.

Code and training · Evaluation reports

Model repository

The detector compares encoder blocks 1, 2 and 3 with the final posterior-mean representation of a frozen FLUX.2 VAE. Each stream is projected to an 8 × 8 token grid. The three early-to-final differences are fused with all four feature streams before transformer classification. Inputs are native-resolution 128 × 128 RGB crops with a central 96 × 96 active region. The VAE decoder and generative FLUX transformer are not used.

The training run excludes CelebA-HQ from training, validation and patch caches; it is used only for the separate internal evaluation below. It contains 21,464 training, 1,257 validation and 1,446 test records. Large starts with newly initialized trainable parameters and the licensed frozen VAE encoder. Students learn from this new Large through temperature-2 Bernoulli KL soft-target distillation: 50% distillation and 50% supervised classification plus paired ranking. Checkpoint selection uses validation recall at a target false-positive rate of 1%, with ROC-AUC as the tie-breaker. A decline receives two additional recovery epochs unless a sustained overfitting signal is observed.

Image-level inference averages five crop logits. Scores are not calibrated probabilities. All benchmark tables below use the internal CAT evaluation and unchanged release thresholds.

Internal benchmark: CAT CelebA-HQ

All 2,000 eligible images in the archived subset: 200 per protection method and 200 clean controls. Noisy baseline excluded. Original five-crop inference and unchanged release thresholds, not experimental sliding-window calibration.

Detection recall

Percentage of protected images correctly detected. Higher is better.

Clean-image specificity

Percentage of clean images correctly accepted as clean. Higher is better. All models use the same 200 clean controls.

Overall results

Bold marks the highest per-method recall or specificity, including ties. Published reference results are listed separately below.

Evaluation-only source: CAT, CelebA-HQ subset. This subset was excluded from training and validation of the currently released checkpoints and was not used to fit their thresholds. The project author confirms having reviewed the source terms and permission for this evaluation use. Images are not redistributed. Protection methods share underlying originals.

Detailed protocol and results.

Published reference alongside the internal benchmark

Detection recall

Clean-image specificity

Our numbers come exclusively from the internal CAT benchmark above. Clean specificity uses the same 200 controls for every row. LightShed numbers are published results on its own evaluation, not measurements on this internal set. Glaze 2 is shown alongside published Glaze; versions and settings are not asserted identical. Bold marks the larger reported value, including ties, not a controlled head-to-head win. Not reported means absent from Table 2. LightShed source, Table 2.

License and data provenance

The author licenses the code and the rights they hold in the new model weights under Apache-2.0. The frozen BFL VAE component is Apache-2.0. This grant does not establish third-party rights in training images. Retained sources include VGGFace2 and WikiArt through CAT, robust-style-mimicry, XAI, Helen, Open Images, DiffVax and user-supplied data; their commercial permissions have not all been verified. See the provenance report.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{ifmain2026datapoisoningdetector,
    title = {DataPoisoning Detector v1: Multi-Level FLUX VAE Features for Protective-Perturbation Detection},
    url = {https://github.com/ifmain/datapoisoning-detector-v1},
    author = {{ifmain}},
    month = {October},
    year = {2026}
}

Selected protection and dataset references: REFERENCES.bib.

Component layout

detector/ contains the trained projection, fusion and classification head. vae/ contains the frozen FLUX VAE encoder and its configuration; no decoder is included. Load the repository root or its detector/ directory with dpdetector.load_detector. The two component state dictionaries reconstruct the original checkpoint exactly.