ArtifactNet: A Lightweight Forensic Framework That Detects AI-Generated Music Using Codec Physics

It starts from the observation that commercial AI music generators like Suno, Udio, MusicGen, and Stable Audio all pass through Residual Vector Quantization (RVQ).

RVQ maps continuous audio into discrete codebook vectors, and the quantization gap produced in this process is irreversible. When a source separation model trained only on human music processes AI-generated audio, this gap manifests as an abnormally large and structured reconstruction residual, which is used as the forensic signal.

Existing detectors like CLAM and SpecTTTra work within their training distribution but easily break down on new generators. ArtifactNet, by contrast, captures not "how it sounds" but "why it is physically different."

The total configuration is 4.0M parameters.

• ArtifactUNet (3.6M): a bounded-mask UNet that predicts a multiplicative mask in the range [0, 0.5] on the STFT magnitude
• 2-stage knowledge distillation: using Demucs v4's residuals as the teacher
• 7-channel HPSS forensic features: decomposing residuals into harmonic/percussive components, combined with time derivatives and spectral flux
• Lightweight CNN (0.4M): processes 4-second segments and makes a track-level judgment via median

As forensic evidence, the effective bandwidth of source separation residuals was measured (n=94).

• Human music: average 1,996 Hz
• AI average (22 generators): 291 Hz
• Suno v3.5: 170 Hz
• Riffusion: 219 Hz
• MusicGen: 255 Hz

Regardless of generator architecture, all AI generators clustered largely around 200 Hz.

ArtifactBench performance is based on 6,183 tracks, 22 generators, with no train-test overlap.

• ArtifactNet: F1 0.983, FPR 1.5%, 4M parameters
• CLAM: F1 0.758, FPR 69.3%, 194M parameters
• SpecTTTra: F1 0.771, FPR 19.4%, 19M parameters

In particular, it points out that CLAM misclassifies real music as AI at a rate of 69.3%, making it essentially unusable as a discriminator. The SONICS/MoM benchmark's real set is centered on YouTube IDs, making original-source comparison difficult due to deletions and privatization, whereas ArtifactBench provides an equal-condition comparison using a directly collected and verified real partition.

The limitations are also clear.

• Requires 44.1kHz input
• FPR of about 8% on low-bitrate MP3
• TPR drops to 94% under a single-pass Demucs laundering attack
• Udio TPR 87% for the latest version

The demo, paper, model, and benchmark have been released together, and the model is distributed under a CC BY-NC 4.0 license. A patent is being filed under KR + PCT.