intrect/artifactnet

1---2license: cc-by-nc-4.03tags:4  - audio-classification5  - ai-music-detection6  - forensic7  - onnx8language:9  - en10pipeline_tag: audio-classification11---12 13# ArtifactNet v9.4 — AI-Generated Music Forensic Detection14 15ArtifactNet detects AI-generated music by extracting forensic residual artifacts via a task-specific UNet, rather than learning generator-specific patterns. This approach generalizes across 22 AI music generators with only 4.2M parameters.16 17**Try it without installing anything:** [analyze a track in the live ArtifactNet demo](https://demo.intrect.io/?utm_source=huggingface&utm_medium=referral&utm_campaign=artifactnet_hub_2026q3&utm_content=model_card). For the production API, batch workflow, limitations, and current plans, see the [ArtifactNet product page](https://intrect.io/artifactnet/?utm_source=huggingface&utm_medium=referral&utm_campaign=artifactnet_hub_2026q3&utm_content=model_card).18 19> ⚠️ **License: CC BY-NC 4.0 — Non-Commercial Only**20> This ONNX inference build may not be used for any commercial product, service, API, or21> revenue-generating activity. Research, academic, and personal evaluation use are welcome.22> For commercial licensing, contact: **contact@intrect.io**23 24> 🛡️ **Patent Pending (KR + PCT)**25> The bounded-mask residual extraction and codec-invariant training methods used in26> ArtifactNet are covered by pending patent applications. Use of the ONNX build under27> CC BY-NC 4.0 grants no patent license; commercial deployment requires both a28> commercial license and a patent license (contact above for both).29 30> ℹ️ **What is released**31> A pre-compiled, end-to-end **ONNX inference build** of the full pipeline (STFT → UNet →32> HPSS → 7-channel CNN → sigmoid). Raw PyTorch weights, training code, and training data33> are **not** publicly released. This is a deliberate scope limitation — the released34> binary is sufficient to reproduce inference numbers reported in our paper, but does35> not enable fine-tuning or weight extraction.36 37## Model Description38 39- **Architecture**: ArtifactUNet (3.6M) + 7ch HPSS Forensic CNN (424K) = 4.2M total40- **Input**: 44.1kHz mono audio, 4-second segments41- **Output**: P(AI) ∈ [0, 1] per segment, song-level median verdict42- **Format**: Single ONNX file (entire pipeline: STFT → UNet → HPSS → 7ch → CNN → sigmoid)43 44## Performance — ArtifactBench v0.9 (test-only fair eval, all models unseen)45 46| Metric | ArtifactNet (4.2M) | CLAM (194M) | SpecTTTra (19M) |47|---|---|---|---|48| **F1** | **0.9829** | 0.7576 | 0.7713 |49| **Precision** | 0.9905 | 0.6674 | 0.8519 |50| **Recall (TPR)** | 0.9755 | 0.8761 | 0.7046 |51| **FPR** | 0.0149 | 0.6926 | 0.1943 |52| **AUC** | **0.9974** | 0.7031 | 0.8460 |53| @FPR≤5% TPR | **99.1%** | - | - |54 55Evaluated on 2,263 tracks (`bench_origin=test`, unseen by all three models),56threshold τ=0.5, identical preprocessing.57 58## Usage59 60```python61import onnxruntime as ort62import numpy as np63import soundfile as sf64 65# Load ONNX inference build66sess = ort.InferenceSession("artifactnet_v94_full.onnx")67 68# Load audio (44.1kHz mono, 4-second chunk)69audio, sr = sf.read("track.wav", dtype="float32")70if audio.ndim > 1:71    audio = audio.mean(axis=1)72chunk = audio[:4 * 44100].reshape(1, -1).astype(np.float32)73 74# Inference75prob = sess.run(None, {"audio": chunk})[0][0]76print(f"P(AI) = {prob:.4f}")  # > 0.5 → AI-generated77```78 79For song-level verdict, compute median over multiple chunks.80 81## Benchmark82 83Evaluate with [ArtifactBench v1](https://huggingface.co/datasets/intrect/artifactbench).84 85If you want to compare the public ONNX build with the production detector on your own audio, use the [live demo](https://demo.intrect.io/?utm_source=huggingface&utm_medium=referral&utm_campaign=artifactnet_hub_2026q3&utm_content=model_benchmark).86 87## Citation88 89```bibtex90@article{oh2026artifactnet,91  title        = {ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics},92  author       = {Oh, Heewon},93  journal      = {arXiv preprint arXiv:2604.16254},94  year         = {2026},95  eprint       = {2604.16254},96  archivePrefix= {arXiv},97  primaryClass = {cs.SD},98  doi          = {10.48550/arXiv.2604.16254},99  url          = {https://arxiv.org/abs/2604.16254}100}101```102 103**arXiv**: [2604.16254](https://arxiv.org/abs/2604.16254) · **DOI**: [10.48550/arXiv.2604.16254](https://doi.org/10.48550/arXiv.2604.16254)104 105## License106 107**CC BY-NC 4.0** — Free for academic, research, and personal use. **Commercial use is108prohibited** without prior written permission. This includes (but is not limited to):109 110- Selling access to the ONNX build or its outputs111- Integrating into commercial products, SaaS, or APIs112- Using the model to generate revenue, directly or indirectly113- Attempting to extract weights for derivative commercial models114 115For commercial licensing inquiries: **contact@intrect.io**116 117### Patent Notice118 119Patent applications covering the core methods of ArtifactNet are pending in Korea (KR)120and via the Patent Cooperation Treaty (PCT). The CC BY-NC 4.0 license on this ONNX121inference build does **not** convey any patent rights. Commercial use, even under a122commercial copyright license, requires a separate patent license. Academic and123research use within the scope of CC BY-NC 4.0 is permitted without separate patent124license, consistent with standard research-use exemptions.125