ArtifactNet v9.4 — AI-Generated Music Forensic Detection

ArtifactNet 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.

⚠️ License: CC BY-NC 4.0 — Non-Commercial Only This ONNX inference build may not be used for any commercial product, service, API, or revenue-generating activity. Research, academic, and personal evaluation use are welcome. For commercial licensing, contact: unohee.official@gmail.com

🛡️ Patent Pending (KR + PCT) The bounded-mask residual extraction and codec-invariant training methods used in ArtifactNet are covered by pending patent applications. Use of the ONNX build under CC BY-NC 4.0 grants no patent license; commercial deployment requires both a commercial license and a patent license (contact above for both).

ℹ️ What is released A pre-compiled, end-to-end ONNX inference build of the full pipeline (STFT → UNet → HPSS → 7-channel CNN → sigmoid). Raw PyTorch weights, training code, and training data are not publicly released. This is a deliberate scope limitation — the released binary is sufficient to reproduce inference numbers reported in our paper, but does not enable fine-tuning or weight extraction.

Model Description

• Architecture: ArtifactUNet (3.6M) + 7ch HPSS Forensic CNN (424K) = 4.2M total
• Input: 44.1kHz mono audio, 4-second segments
• Output: P(AI) ∈ [0, 1] per segment, song-level median verdict
• Format: Single ONNX file (entire pipeline: STFT → UNet → HPSS → 7ch → CNN → sigmoid)

Performance — ArtifactBench v0.9 (test-only fair eval, all models unseen)

Metric	ArtifactNet (4.2M)	CLAM (194M)	SpecTTTra (19M)
F1	0.9829	0.7576	0.7713
Precision	0.9905	0.6674	0.8519
Recall (TPR)	0.9755	0.8761	0.7046
FPR	0.0149	0.6926	0.1943
AUC	0.9974	0.7031	0.8460
@FPR≤5% TPR	99.1%	-	-

Evaluated on 2,263 tracks (bench_origin=test, unseen by all three models), threshold τ=0.5, identical preprocessing.

Usage

import onnxruntime as ort
import numpy as np
import soundfile as sf

# Load ONNX inference build
sess = ort.InferenceSession("artifactnet_v94_full.onnx")

# Load audio (44.1kHz mono, 4-second chunk)
audio, sr = sf.read("track.wav", dtype="float32")
if audio.ndim > 1:
    audio = audio.mean(axis=1)
chunk = audio[:4 * 44100].reshape(1, -1).astype(np.float32)

# Inference
prob = sess.run(None, {"audio": chunk})[0][0]
print(f"P(AI) = {prob:.4f}")  # > 0.5 → AI-generated

For song-level verdict, compute median over multiple chunks.

Benchmark

Evaluate with ArtifactBench v1.

Citation

@article{oh2026artifactnet,
  title={ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics},
  author={Oh, Heewon},
  year={2026}
}

License

CC BY-NC 4.0 — Free for academic, research, and personal use. Commercial use is prohibited without prior written permission. This includes (but is not limited to):

• Selling access to the ONNX build or its outputs
• Integrating into commercial products, SaaS, or APIs
• Using the model to generate revenue, directly or indirectly
• Attempting to extract weights for derivative commercial models

For commercial licensing inquiries: unohee.official@gmail.com

Patent Notice

Patent applications covering the core methods of ArtifactNet are pending in Korea (KR) and via the Patent Cooperation Treaty (PCT). The CC BY-NC 4.0 license on this ONNX inference build does not convey any patent rights. Commercial use, even under a commercial copyright license, requires a separate patent license. Academic and research use within the scope of CC BY-NC 4.0 is permitted without separate patent license, consistent with standard research-use exemptions.