Add UTCN-XLSR2B-LogReg submission for LRLspoof (1-SRR=7.2818)

#38
by korallll - opened
Files changed (1) hide show
  1. submissions/utcn-xlsr2b-logreg.yaml +73 -0
submissions/utcn-xlsr2b-logreg.yaml ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: UTCN-XLSR2B-LogReg
4
+ slug: utcn-xlsr2b-logreg
5
+ description: 'Data-centric real-world deepfake detector (Combei et al., Interspeech 2025): frozen wav2vec
6
+ 2.0 XLS-R 2B truncated to its first 9 transformer layers (the authors'' released "wav2vec2-xls-r-2b_truncated"),
7
+ features = hidden_states[9] of the truncated model (= output of layer 9 after the encoder''s final
8
+ LayerNorm, as in the authors'' app.py), mean-pooled over time (1920-d), scored by a logistic regression
9
+ (C=1e6) trained on the margin-pruned union of seven scientific corpora ("ALL": ASVspoof 2019 LA, FoR-norm,
10
+ ASVspoof 2021 subset, TIMIT-TTS, ODSS, MLAAD v5 + M-AILABS, ASVspoof 5 train/dev), followed by the
11
+ authors'' MinMax score scaler. Input: 16 kHz, Arena 4 s cap (first 64600 samples, shorter clips repeat-padded;
12
+ app.py scores <= 60 s segments), per-utterance zero-mean/unit-variance normalisation as the HF feature
13
+ extractor. score = MinMax-scaled LR decision value (class 1 = real), higher = bona fide. FP32 (TF32
14
+ off, eager attention). Authors'' released checkpoint; no license, so the weights are linked, not re-hosted.'
15
+ code: https://huggingface.co/DavidCombei/deepfake-detection/blob/5f39744bc81762529e19f76e2bb7318e7e69f3fb/audio-detection_core/app.py
16
+ checkpoint: https://huggingface.co/DavidCombei/deepfake-detection/tree/5f39744bc81762529e19f76e2bb7318e7e69f3fb/audio-detection_core/model
17
+ params_millions: 433.051009
18
+ paper:
19
+ arxiv_id: '2506.09606'
20
+ url: https://arxiv.org/abs/2506.09606
21
+ bibtex: "@inproceedings{combei2025unmasking,\n title = {Unmasking real-world audio deepfakes:\
22
+ \ A data-centric approach},\n author = {Combei, David and Stan, Adriana and Oneata, Dan and\
23
+ \ M{\\\"u}ller, Nicolas and Cucu, Horia},\n booktitle = {Proc. Interspeech 2025},\n year \
24
+ \ = {2025},\n note = {arXiv:2506.09606}\n}\n"
25
+ dataset:
26
+ id: lab260/LRLspoof
27
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
28
+ split: test
29
+ scores:
30
+ srr_complement: 7.281801668342059
31
+ n_trials: 1304455
32
+ n_skipped: 286
33
+ calibration:
34
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
35
+ threshold: 0.573857757173956
36
+ artifact:
37
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/UTCN-XLSR2B-LogReg/resolve/a21270058c575189986c90e833b07b28b9be690e/.eval_results/lab260/LRLspoof/scores.txt
38
+ scores_sha256: a5dfdfc5e52666e739fcf7384a327ee0062ff1a0a65d5f88a44a2ef03109074a
39
+ bench_version: speech-spoof-bench==0.4.1
40
+ reproduction:
41
+ reproduced_by: SpeechAntiSpoofingBenchmarks
42
+ reproduced_at: '2026-11-15'
43
+ reproduced_bench_version: speech-spoof-bench==0.4.1
44
+ match: inference
45
+ submitter:
46
+ hf_username: korallll
47
+ contact: k.n.borodin@mtuci.ru
48
+ submitted_at: '2026-11-15'
49
+ notes: 'UTCN-XLSR2B-LogReg = HF DavidCombei/deepfake-detection @ 5f39744bc81762529e19f76e2bb7318e7e69f3fb,
50
+ audio core only: app.py (inference), model/wav2vec2-xls-r-2b_truncated/model.safetensors (182 fp32 tensors,
51
+ 433,051,008 params; strict load into a 9-layer transformers Wav2Vec2Model, 0 missing / 0 unexpected)
52
+ and model/logreg_margin_pruning_ALL_with_scaler_threshold.joblib (sha256 dbdc1bd864d2b997fc873499f6485891998d5f829af4746064426e942ba580dc).
53
+ The joblib is a pickle of (sklearn 1.3.2 LogisticRegression, MinMaxScaler, threshold 0.56713); it was
54
+ never joblib.load-ed: an allow-list unpickler (globals: LogisticRegression, MinMaxScaler, joblib NumpyArrayWrapper,
55
+ numpy.ndarray/dtype only, all mapped to inert stubs) exported coef/intercept/scaler/threshold to lr_head.npz
56
+ (sha256 044bbb7e6f961a7871007414bdc70ebb99500e602b0b93c53b267448c7ff6684), which is all the wrapper
57
+ reads. No license on the repo -> weights linked, not re-hosted. Feature path exactly as app.py (normalisation,
58
+ hidden_states[9] of the truncated model incl. the encoder''s final LayerNorm, mean pool, decision_function,
59
+ MinMax transform); verified against the authors'' path (app.py model class + sklearn objects rebuilt
60
+ from the npz) on 18 J-SPAW_LA clips: max |d| 0 (whole clip) / 2.5e-8 (4 s cap) with eager attention;
61
+ 1.2e-3 vs the transformers-default SDPA attention, which is itself run-to-run non-deterministic in FP32
62
+ here (the wrapper uses eager attention for bit-reproducibility). Deviations: Arena 4 s cap (first 64600
63
+ samples, repeat-pad) instead of app.py''s 60 s segments averaged; the score is the per-window scaled
64
+ decision value (app.py thresholds it at 0.56213 and reports 1 - mean sigmoid). Training (paper sec.
65
+ 2.1, 4.1, 4.2): all samples incl. test splits of ASVspoof 2019 LA, FoR (for-norm), ~60k random ASVspoof
66
+ 2021 utterances (22k bona fide / 589k spoof pool = the DF eval set), TIMIT-TTS, ODSS, MLAAD v5 + 80k
67
+ M-AILABS bona fide, ASVspoof 5 train+dev; margin-based sample pruning. In-the-Wild and AI4T were not
68
+ trained on, but the released configuration (ALL + margin pruning) is the one selected by its In-the-Wild
69
+ EER; selection only, so InTheWild is ranked normally (maintainer rule 2026-11-07, as for XLSR-Nes2Net-LA).
70
+ ASVspoof2019_LA, ASVspoof2021_DF and ODSS marked (DECRO follows from ASVspoof2019 LA eval via provenance).
71
+ Self-reported EER% (paper Table 4, ALL, margin pruning, no augmentation): In-the-Wild 1.70, AI4T 12.43.
72
+ J-SPAW_LA EER 0.875% with this wrapper (PyTorch; FP32 TensorRT engines on RTX 4070 Ti SUPER and RTX
73
+ 5060 Ti give the same EER, Spearman 1.000000, max |d| 1e-6).'