intrect/artifactbench
Viewer • Updated • 4.4k • 1.87k
Heewon Oh
intrect
AI & ML interests
NLP, MIR, LLM finetune, Quant
Recent Activity
updated a dataset 12 days ago
intrect/artifactbench updated a Space 21 days ago
intrect/artifactnet updated a model 21 days ago
intrect/artifactnetOrganizations
updated a
dataset 12 days ago
updated a
Space 21 days ago
updated a
model 21 days ago
posted an update 6 months ago
Post
217
I’m excited to share a new paper I recently posted on arXiv: ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics.
This work asks a simple question: can AI-generated music be detected not only by style, but by the physical artifacts left behind during generation?
ArtifactNet approaches the problem from that angle. Instead of only learning what AI music sounds like on a fixed benchmark, it analyzes forensic residual patterns linked to neural audio codec bottlenecks such as residual vector quantization (RVQ).
In our experiments, ArtifactNet achieved F1 = 0.9829 on a zero-overlap multi-generator benchmark spanning 22 AI generators and 6 real-music sources, while using only 4.0M parameters. Under the same evaluation setting, larger prior models showed substantial degradation on out-of-distribution generators and real-music false positives.
I also introduced ArtifactBench, a broader evaluation benchmark designed to stress-test detector robustness across unseen generators, diverse real sources, hard negatives, and codec conditions.
This was a deeply rewarding project at the intersection of audio forensics, MIR, and generative model evaluation.
https://arxiv.org/abs/2604.16254
This work asks a simple question: can AI-generated music be detected not only by style, but by the physical artifacts left behind during generation?
ArtifactNet approaches the problem from that angle. Instead of only learning what AI music sounds like on a fixed benchmark, it analyzes forensic residual patterns linked to neural audio codec bottlenecks such as residual vector quantization (RVQ).
In our experiments, ArtifactNet achieved F1 = 0.9829 on a zero-overlap multi-generator benchmark spanning 22 AI generators and 6 real-music sources, while using only 4.0M parameters. Under the same evaluation setting, larger prior models showed substantial degradation on out-of-distribution generators and real-music false positives.
I also introduced ArtifactBench, a broader evaluation benchmark designed to stress-test detector robustness across unseen generators, diverse real sources, hard negatives, and codec conditions.
This was a deeply rewarding project at the intersection of audio forensics, MIR, and generative model evaluation.
https://arxiv.org/abs/2604.16254
- 1 reply
submitted a
paper to Daily Papers 6 months ago
[bot] Conversion to Parquet
#1 opened 6 months ago
by 
parquet-converter
published a
dataset 6 months ago
published a
model 6 months ago
vela
#1 opened 6 months ago
by 
willowoods
upvoted a paper 6 months ago
updated a
model 6 months ago
updated a
dataset 6 months ago
published a
dataset 6 months ago
updated a
Space 6 months ago
published a
Space 6 months ago
published a
Space 8 months ago
updated 2
Spaces 8 months ago
updated a
model 8 months ago
published a
Space 8 months ago