Overview
This research introduces a method for detecting AI-generated music that maintains performance under audio manipulations such as speed modification and pitch shifting. The approach addresses a robustness problem in existing detectors, whose efficacy diminishes significantly when confronted with altered audio. The proposed detection pipeline is designed to be frequency-scaling-invariant, providing resilience against such adversarial modifications.
Research Context
AI music generators produce musical tracks that contain predictable spectral artifacts, which are determined by their underlying architectural design. Prior detection mechanisms have been developed to exploit these artifacts, achieving high accuracy when analyzing raw, unmanipulated generated tracks. However, the effectiveness of these detectors collapses following simple audio manipulations, specifically speed modification or pitch shifting. This vulnerability represents an open problem in the field of AI-generated content detection.
Approach
The core of the presented method is a frequency-scaling-invariant detection pipeline. This pipeline prevents performance collapse under speed modification or pitch shifting by design. The methodology involves mapping audio data onto a log-frequency axis. This mapping is achieved through a log-Short-Time Fourier Transform (log-STFT) remapping process. Following this transformation, a single learned cross-correlation filter is applied. This filter, in conjunction with max-pooling, ensures shift invariance during the inference phase of the detection process.
The training regimen for this system employs a hybrid loss function. This loss function jointly supervises two objectives: binary detection and artifact-peak localization. The inclusion of artifact-peak localization serves to regularize the boundary weights within the model. A direct consequence of building robustness to speed changes into the design is that the detector also yields interpretable outputs. Beyond a binary decision regarding the AI-generated nature of the music, the system provides an estimate of any applied speed-change factor.
Findings
- AI music generators embed predictable spectral artifacts linked to their architecture.
- Existing AI music detectors demonstrate high accuracy on raw generated tracks.
- Performance of existing detectors degrades significantly under simple audio manipulations like speed modification or pitch shifting.
- The proposed frequency-scaling-invariant detection pipeline aims to prevent performance collapse under these manipulations.
- The method maps audio onto a log-frequency axis via log-STFT remapping.
- A learned cross-correlation filter combined with max-pooling provides shift invariance at inference.
- Training utilizes a hybrid loss function that supervises both binary detection and artifact-peak localization, which regularizes boundary weights.
- The detector provides a binary decision on AI generation and an estimate of the applied speed-change factor, stemming from its built-in robustness to speed alteration.
Potential Applications
The built-in robustness to speed change, coupled with the detector's interpretability, suggests potential use cases where identifying audio manipulation or verifying content origin despite such modifications is critical. The system's ability to output an estimated speed-change factor alongside its binary decision offers additional diagnostic information.
Key Limitations Mentioned by Researchers
The Abstract does not explicitly mention limitations of the presented research.