Improved Robustness in AI-Generated Music Detection

2026-07-29

AI Summary

This work addresses the significant drop in robustness of existing AI-generated music detectors when confronted with simple audio transformations such as tempo changes and pitch shifts. To overcome this limitation, the authors propose a novel detection architecture that inherently incorporates frequency scaling invariance. The approach maps audio signals onto a logarithmic frequency axis via log-STFT and combines learnable cross-correlation filters with max-pooling to achieve translation invariance during inference. This is the first method to integrate frequency scaling invariance directly into the detection pipeline, simultaneously producing both a binary authenticity decision and an estimate of the applied tempo scaling factor, thereby enhancing model interpretability and adversarial robustness. Experimental results demonstrate that the proposed method maintains high detection accuracy under various audio transformation attacks, substantially outperforming current state-of-the-art techniques while accurately estimating transformation parameters.

Abstract

AI music generators leave predictable spectral artifacts determined by their architecture. Existing detectors exploit these artifacts with near-perfect accuracy on raw generated tracks, but their performance collapses under simple audio manipulations, such as speed modification or pitch shifting. We address this open robustness problem by introducing a frequency-scaling-invariant detection pipeline that aims to prevent this kind of attack by design. Our method maps audio onto a log-frequency axis via a log-STFT remapping. A single learned cross-correlation filter, combined with max-pooling, provides shift invariance at inference time. Training uses a hybrid loss that jointly supervises binary detection and artifact-peak localization, regularizing boundary weights. Because robustness to speed change is built in by design, the detector is also interpretable: it outputs both a binary decision and an estimate of the applied speed-change factor.

Problem

• AI-generated music detection
• robustness
• audio manipulation
• speed modification
• pitch shifting

Innovation

• frequency-scaling-invariant
• log-STFT remapping
• cross-correlation filter
• hybrid loss
• artifact localization