partial-spoof-mrm-audit

MRM (Multi-Resolution Model): Partial Spoof Detector Checkpoint

Checkpoint accompanying the paper "How Trustworthy Are Partial Spoof Detectors? A Cross-Domain Operational Audit", accepted at IJCB 2026, Special Session 1: Trustworthy and Secure AI for Behavioural and Biometric Recognition. Yash Sukhdeve, Ajan Ahmed, and Masudul H. Imtiaz, Department of Electrical and Computer Engineering, Clarkson University.

Analysis code and verification suite: https://github.com/AVHBAC/partial-spoof-cross-domain-audit-reproducibility

What this is

A multi-resolution partial-spoof detector: the model of Zhang et al. (the PartialSpoof multi-resolution countermeasure), via the open reimplementation MultiResoModel-Simple (Luong et al.), trained by us on the PartialSpoof training set. Front-end: wav2vec 2.0 Large; back-end: losses supervised jointly at frame (20 ms), segment, and utterance scales.

This is not an authors'-released checkpoint of the original model, and not the public MultiResoModel-Simple checkpoint. It is our own training run.

Note that authors' released multi-resolution checkpoints do exist: the PartialSpoof repository ships 03multireso/01_download_pretrained_models.sh, which retrieves multi-reso.tar.gz from Zenodo record 6674660. This checkpoint is an independent retraining, not a substitute for those artifacts, and results obtained with it should not be read as reproducing them.

File

Reported performance (PartialSpoof eval, ours)

• Utterance-level EER: 0.94%
• Segment-level (20 ms frame) EER: 13.91%

For reference, the public reimplementation checkpoint reports ~1.48% / ~13.67%, and the original Zhang et al. model reports 0.49% utterance-level EER. Cross-domain behaviour (LlamaPartialSpoof, PartialEdit, HQ-MPSD) is the subject of the paper.

These thresholds and scores are derived from the PartialSpoof evaluation split; no separate development split was held out. In-domain figures are therefore optimistic.

Training recipe

• Multi-resolution units {0.02, 0.04, 0.08, 0.16, 0.32, 0.64} s; segment duration 9.6 s; random_seek = true, use_mask = true.
• Optimizer Adam, lr 1e-5; scheduler StepLR, step 10, gamma 0.5; loss P2SGradLoss.
• Batch size 8, 6 dataloader workers. Trained to epoch 55.
• Implementation: github.com/hieuthi/MultiResoModel-Simple @ 0f69db3a2d654de47822d951fe6ad256bbaac9ba.

Random seeding

A fixed seed was used. train.py defaults to --seed 1234 and calls reproducibility(seed), which seeds torch, random, numpy, PYTHONHASHSEED and CUDA, and sets cudnn.deterministic = True, cudnn.benchmark = False. The launch script passes no override, so the default applies. The random_seek crop draws from Python's random, which PyTorch seeds deterministically per dataloader worker.

Earlier revisions of this card stated that no seed was fixed. That was incorrect and is corrected here.

Bit-identical reproduction is still not guaranteed, for two remaining reasons: torch.use_deterministic_algorithms(True) is not set, so some CUDA kernels used during wav2vec 2.0 fine-tuning may vary run to run; and the original training log was not retained, so the seed in force can be established from the committed script but not from a run record. The published weights, not the recipe, are the authoritative artifact behind every MRM number in the paper.

Intended use

Research and reproducibility only. This checkpoint supports an audit of an existing detector design under cross-domain partial-spoof attacks. It is not a deployable forensic tool, and no claim is made about its reliability outside the corpora studied.

Related artifacts

• Frame-level and utterance-level detector score outputs: huggingface.co/datasets/sukhdeveyash/partial-spoof-cross-domain-audit-data

License & attribution

Released under MIT, following the MultiResoModel-Simple reimplementation (MIT). If you use this checkpoint, please cite the original multi-resolution model (Zhang et al., IEEE/ACM TASLP 2023, The PartialSpoof Database and Countermeasures for the Detection of Short Fake Speech Segments Embedded in an Utterance) and the reimplementation (Luong et al., ICASSP 2025, LlamaPartialSpoof), together with the paper above.