Add UTCN-XLSR2B-LogReg submission for DECRO (EER=6.3370)

submissions/utcn-xlsr2b-logreg.yaml

@@ -0,0 +1,70 @@
+schema_version: 4
+system:
+  name: UTCN-XLSR2B-LogReg
+  slug: utcn-xlsr2b-logreg
+  description: 'Data-centric real-world deepfake detector (Combei et al., Interspeech 2025): frozen wav2vec
+    2.0 XLS-R 2B truncated to its first 9 transformer layers (the authors'' released "wav2vec2-xls-r-2b_truncated"),
+    features = hidden_states[9] of the truncated model (= output of layer 9 after the encoder''s final
+    LayerNorm, as in the authors'' app.py), mean-pooled over time (1920-d), scored by a logistic regression
+    (C=1e6) trained on the margin-pruned union of seven scientific corpora ("ALL": ASVspoof 2019 LA, FoR-norm,
+    ASVspoof 2021 subset, TIMIT-TTS, ODSS, MLAAD v5 + M-AILABS, ASVspoof 5 train/dev), followed by the
+    authors'' MinMax score scaler. Input: 16 kHz, Arena 4 s cap (first 64600 samples, shorter clips repeat-padded;
+    app.py scores <= 60 s segments), per-utterance zero-mean/unit-variance normalisation as the HF feature
+    extractor. score = MinMax-scaled LR decision value (class 1 = real), higher = bona fide. FP32 (TF32
+    off, eager attention). Authors'' released checkpoint; no license, so the weights are linked, not re-hosted.'
+  code: https://huggingface.co/DavidCombei/deepfake-detection/blob/5f39744bc81762529e19f76e2bb7318e7e69f3fb/audio-detection_core/app.py
+  checkpoint: https://huggingface.co/DavidCombei/deepfake-detection/tree/5f39744bc81762529e19f76e2bb7318e7e69f3fb/audio-detection_core/model
+  params_millions: 433.051009
+  paper:
+    arxiv_id: '2506.09606'
+    url: https://arxiv.org/abs/2506.09606
+    bibtex: "@inproceedings{combei2025unmasking,\n  title     = {Unmasking real-world audio deepfakes:\
+      \ A data-centric approach},\n  author    = {Combei, David and Stan, Adriana and Oneata, Dan and\
+      \ M{\\\"u}ller, Nicolas and Cucu, Horia},\n  booktitle = {Proc. Interspeech 2025},\n  year     \
+      \ = {2025},\n  note      = {arXiv:2506.09606}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 6.337013922227556
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/UTCN-XLSR2B-LogReg/resolve/84d1b5117ea1f2acd77b050dd6c8c46276e36b88/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 17eec593c9de864085e851d0568fcfa44afd2e7b3a727c8f77096fbee674deac
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-10'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-10'
+notes: 'UTCN-XLSR2B-LogReg = HF DavidCombei/deepfake-detection @ 5f39744bc81762529e19f76e2bb7318e7e69f3fb,
+  audio core only: app.py (inference), model/wav2vec2-xls-r-2b_truncated/model.safetensors (182 fp32 tensors,
+  433,051,008 params; strict load into a 9-layer transformers Wav2Vec2Model, 0 missing / 0 unexpected)
+  and model/logreg_margin_pruning_ALL_with_scaler_threshold.joblib (sha256 dbdc1bd864d2b997fc873499f6485891998d5f829af4746064426e942ba580dc).
+  The joblib is a pickle of (sklearn 1.3.2 LogisticRegression, MinMaxScaler, threshold 0.56713); it was
+  never joblib.load-ed: an allow-list unpickler (globals: LogisticRegression, MinMaxScaler, joblib NumpyArrayWrapper,
+  numpy.ndarray/dtype only, all mapped to inert stubs) exported coef/intercept/scaler/threshold to lr_head.npz
+  (sha256 044bbb7e6f961a7871007414bdc70ebb99500e602b0b93c53b267448c7ff6684), which is all the wrapper
+  reads. No license on the repo -> weights linked, not re-hosted. Feature path exactly as app.py (normalisation,
+  hidden_states[9] of the truncated model incl. the encoder''s final LayerNorm, mean pool, decision_function,
+  MinMax transform); verified against the authors'' path (app.py model class + sklearn objects rebuilt
+  from the npz) on 18 J-SPAW_LA clips: max |d| 0 (whole clip) / 2.5e-8 (4 s cap) with eager attention;
+  1.2e-3 vs the transformers-default SDPA attention, which is itself run-to-run non-deterministic in FP32
+  here (the wrapper uses eager attention for bit-reproducibility). Deviations: Arena 4 s cap (first 64600
+  samples, repeat-pad) instead of app.py''s 60 s segments averaged; the score is the per-window scaled
+  decision value (app.py thresholds it at 0.56213 and reports 1 - mean sigmoid). Training (paper sec.
+  2.1, 4.1, 4.2): all samples incl. test splits of ASVspoof 2019 LA, FoR (for-norm), ~60k random ASVspoof
+  2021 utterances (22k bona fide / 589k spoof pool = the DF eval set), TIMIT-TTS, ODSS, MLAAD v5 + 80k
+  M-AILABS bona fide, ASVspoof 5 train+dev; margin-based sample pruning. In-the-Wild and AI4T were not
+  trained on, but the released configuration (ALL + margin pruning) is the one selected by its In-the-Wild
+  EER; selection only, so InTheWild is ranked normally (maintainer rule 2026-11-07, as for XLSR-Nes2Net-LA).
+  ASVspoof2019_LA, ASVspoof2021_DF and ODSS marked (DECRO follows from ASVspoof2019 LA eval via provenance).
+  Self-reported EER% (paper Table 4, ALL, margin pruning, no augmentation): In-the-Wild 1.70, AI4T 12.43.
+  J-SPAW_LA EER 0.875% with this wrapper (PyTorch; FP32 TensorRT engines on RTX 4070 Ti SUPER and RTX
+  5060 Ti give the same EER, Spearman 1.000000, max |d| 1e-6).'