FakeMark: Deepfake Speech Attribution with Watermarked Artifacts

Official pretrained checkpoints for FakeMark, a deepfake speech attribution system. FakeMark injects system-specific watermark artifacts into synthesized speech to attribute waveforms back to their originating Text-to-Speech (TTS) architecture.

---

📌 Overview

FakeMark provides a robust framework for speech provenance. By leveraging SEANet-based generators and MMS-300M collaborators, it allows for high-fidelity audio watermarking that survives common distortions while maintaining high attribution accuracy.

Key Features:

• FakeMarkA: SEANet-based watermark generator with ColProcessor conditioning.
• FakeMarkT: Timbre-based watermark generator.
• Multi-System Support: Includes retrained checkpoints for AudioSeal and Timbre-4bit for benchmarking.

---

📂 Checkpoint Layout

The repository is organized as follows:

checkpoints/
├── FakeMarkA/
│   ├── encoder.ckpt          # SEANet watermark generator
│   ├── decoder.ckpt          # SEANet decoder
│   ├── colprocessor.ckpt     # ColProcessor conditioning module
│   └── collaborator.ckpt     # MMS-300M collaborator (attribution classifier)
├── FakeMarkT/
│   ├── encoder.ckpt          # Timbre watermark generator
│   └── collaborator.ckpt     # MMS-300M collaborator
├── AudioSeal/
│   ├── checkpoint_generator_epoch260.pth   # AudioSeal generator (retrained)
│   └── checkpoint_detector_epoch260.pth    # AudioSeal detector (retrained)
├── Timbre-4bit.pth.tar       # Timbre generator/detector (retrained)
├── MMS_300M.ckpt             # Standalone MMS-300M classifier
└── ResNet.ckpt               # Standalone ResNet34 + LFB + LMCL classifier