Watermark shortcuts in speech deepfake detection: code and results

Code and result files for the paper "Watermark shortcuts in speech deepfake detection: Framing, mitigation and explanation faithfulness". The study trains spoofing countermeasures (frozen XLS-R and WavLM, fine-tuned XLS-R, and a LoRA-adapted Qwen2.5-Omni-3B audio language model) on ASVspoof 2019 LA with spoofs carrying one of four open-source watermarks (AudioSeal, PerTh, SilentCipher, WavMark), and measures framing of watermarked genuine speech, fix transfer, label-blind augmentation, a representation-based risk score, and the faithfulness of the audio language model's explanations. The code archive contains the experiment scripts, and scripts that recompute every table and figure of the paper from the result archive. The result archive contains exactly the files the paper was computed from, with a README and SHA-256 checksums. Speech audio is not redistributed: ASVspoof 2019 LA (doi:10.7488/ds/2555) and In-the-Wild are available from their distributors, and the watermarked copies can be regenerated with the code.