DeepSafe Ensemble Artifacts

The meta-learners that turn 19 individual detector scores into one calibrated verdict, for DeepSafe.

Unlike the model code and weights in the other DeepSafe repositories, these are first-party: trained by us, licensed PolyForm Noncommercial 1.0.0, same as the project.

Contents

Modality	Meta-learner	Held-out AUC
Image	LightGBM over 7 models	0.9466
Audio	Random Forest over 3 models	0.8290
Video	XGBoost over 9 models	0.6694

Each modality ships four files:

• <modality>_meta_learner.pkl — the trained model
• <modality>_scaler.pkl — feature scaling
• <modality>_calibrator.pkl — Platt calibration, so scores read as probabilities
• <modality>_config.json — feature order and fallback weights

Trained on the 15,499-sample medium evaluation tier. Without these, the inference server produces per-model scores but no ensemble verdict.

Usage

setup.sh fetches these automatically. Manually:

from huggingface_hub import snapshot_download
snapshot_download("deepsafe/ensemble", local_dir="models/ensemble/artifacts")

The numbers are the point

Held-out video AUC is 0.6694. That is barely above chance on generators the models were not trained for, and it is lower than the cross-validated figure produced during training. We publish the held-out number because the gap between the two is the finding. See BENCHMARK.md.

Security note

These are Python pickles, which execute code on load. Only load them from a source you trust. DeepSafe's loader refuses any pickle outside its configured artifacts directory, but that is a guardrail, not a guarantee. If you are security-sensitive, retrain your own with deepsafe fit --tier 1.