artifactbench / v1.1 /RESULTS_8WAY.md
intrect's picture
v1.1: add 8-way public model comparison results
7d5482c verified
|
Raw
History Blame Contribute Delete
4.39 kB

ArtifactBench — 8-Way Public Model Comparison (2026-07-04)

Extends the original 3-way comparison (ArtifactNet / SpecTTTra / CLAM) with five additional publicly-available AI-music detection models, evaluated on the same v1.1 purged test partition (n = 2,224; see v1.1/RESULTS_v1.1.md).

Models added

Model Params Source
Deezer ISMIR fakeprint LR 3.6K lofcz/ai-music-detector (Afchar et al., ISMIR 2025)
FST (Mippia) 174.4M Mippia/FST-AI-music-detection
AI-Music-Detection AST-60s 90.8M AI-Music-Detection/ai_music_detection_large_60s
SpecTTTra β-5s 18.7M awsaf49/sonics-spectttra-beta-5s (5s-segment variant of the SONICS α-120s baseline)
DeepFense EAT+Nes2Net (FakeMusicCaps) DeepFense/FakeMusicCaps_EAT_Nes2Net_NoAug_Seed42 (Kheir et al., arXiv:2604.08450)

Adapters: artifactbench/models/{ast_60s,deepfense,deezer_ismir,fst,spectttra_variants}.py. Run with --model ast_60s --model deepfense --model deezer_ismir --model fst --model spectttra_beta5s alongside the existing --model artifactnet --model spectttra --model clam.

Results (n = 2,224, τ = 0.5)

Rank Model Params F1 Precision Recall (TPR)
1 ArtifactNet v9.4 (ours) 4.2M 0.952 0.932 97.3%
2 AI-Music-Detection AST-60s 90.8M 0.840 0.848 83.1%
3 CLAM (MoM) 194.3M 0.787 0.711 88.3%
4 SpecTTTra α-120s 18.7M 0.777 0.880 69.5%
5 Deezer ISMIR fakeprint LR 3.6K 0.754 0.906 64.6%
6 FST (Mippia) 174.4M 0.735 0.984 58.7%
7 DeepFense EAT+Nes2Net 0.650 0.589 72.4%
8 SpecTTTra β-5s 18.7M 0.563 0.884 41.3%

Observation. Parameter count does not predict performance: FST (174.4M) scores below Deezer's 3.6K-parameter logistic-regression baseline. ArtifactNet (4.2M) ranks first among all eight models, including three baselines 20–46× larger.

DeepFense catastrophic-OOD case study. DeepFense (trained on FakeMusicCaps, which draws its AI examples from a different generator/format mix) collapses to near-chance on the SONICS synthetic subset specifically (Chirp/Udio, TPR 0–15%) while performing normally on every other AI-generator subset. This is a genuine distribution-shift failure, not an adapter bug (verified by inspecting per-source means, confirming the split correlates with data source rather than ground-truth label — see the investigation note in this repo's commit history).

Segment-length ablation (SpecTTTra). The same SpecTTTra architecture drops from F1 = 0.777 (120s context) to F1 = 0.563 (5s context) purely from shorter input — evidence that long-context modeling, not just architecture, drives this baseline's detection accuracy.

Reproducibility note — ONNX Runtime CUDA determinism

An earlier run of this 8-way comparison (2026-07-03) produced ArtifactNet F1 = 0.869 using default onnxruntime-gpu settings (cudnn_conv_algo_search: EXHAUSTIVE, the ORT default). Re-running the identical files and code with HEURISTIC algorithm selection (deterministic, avoids per-run algorithm auto-tuning) recovered F1 = 0.952 on the same data. We verified bit-exact reproducibility: two independent runs of a 40-track subset, under both the CPU path and the fixed GPU path, produced identical probability values (max abs diff = 0.0) across runs. The artifactnet.py adapter now pins these settings by default. If you fork this runner and see run-to-run drift with CUDAExecutionProvider on other hardware, check this setting first.

Caveat — file provenance

The real-track set used in this specific 8-way run is a locally-restored superset (some files re-sourced after original storage loss) rather than the exact byte-identical files behind the official v1/v1.1 Table 6 numbers. Absolute FPR values here should be treated as approximate; TPR and cross-model ranking are unaffected (all eight models were scored on identical files). Canonical-byte re-verification is planned for a future update.