From Supra to Sub: Hierarchical Audio Deepfake Source Tracing

Paper: From Supra to Sub: A Hierarchical Margin-Based Approach for Audio Deepfake Source Tracing
Venue: CVPR 2026 APAI Workshop
Authors: Dawid Wolkiewicz, Piotr Syga (Wrocław University of Science and Technology)

Overview

Binary "real vs. fake" detection is well-studied; source tracing — identifying which generative model produced a particular deepfake — is not. This repo implements a two-stage hierarchical attribution framework that first predicts the TTS architecture family (e.g. Tacotron, VITS, XTTS) and then attributes the specific model within that family, with Mahalanobis-distance OOD scoring on top for previously unseen generators.

Getting started

The how-tos live in docs/:

• docs/SETUP.md — clone, uv sync, fetch the MLAAD dataset, generate pre-encoded features
• docs/TRAINING.md — train Flat / H-Shared / H-Arch variants, key arguments
• docs/EVALUATION.md — in-domain classification metrics and OOD detection

The makefile wraps the standard invocations.

Repository layout

.
├── src/adar/             # Installable package (models, datasets, losses, OOD detectors, utils)
├── scripts/              # Runnable entrypoints (prepare, training/, eval, ablation drivers)
├── data/                 # Canonical label / superclass LUTs; MLAAD dataset
├── exp/                  # Training outputs (model checkpoints, logs)
├── docs/                 # Documentation
└── makefile              # Standard pipeline targets

Citation

@InProceedings{Wolkiewicz_2026_CVPR,
    author    = {Wolkiewicz, Dawid and Syga, Piotr},
    title     = {From Supra to Sub: A Hierarchical Margin-Based Approach for Audio Deepfake Source Tracing},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {10770-10778}
}

License

MIT — see LICENSE.