Skip to main content

Beyond Vocals: Harmony-Aware Graph Fusion With Multi-View Features for Music Deepfake Detection

By
Zhaolin Wei; Haoran Duan; Dengpan Ye; Jiacheng Deng; Ziyi Liu; Yueyun Shang; Zhihong Tian; Yanjiao Chen

Music deepfake detection has attracted increasing attention as AI-generated songs pose growing threats to artistic authenticity and copyright protection. Existing audio spoofing detectors primarily focus on vocal analysis, overlooking the complementary forgery cues embedded in background music and pitch dynamics. To address this limitation, we propose a Harmony-Aware Multi-View Graph Fusion (HMGF) framework that jointly models forgery patterns from vocals, background music (BGM), and pitch. The proposed Hybrid Frontend introduces AdaBEATS, an adapter-enhanced variant of the BEATS model that effectively extracts residual forgery artifacts from BGM. Meanwhile, pitch trajectories are analyzed as temporal anomaly sequences using MiniRocket, which captures abrupt pitch transitions in synthesized singing. In the Graph-based Logical Fusion Backend, heterogeneous graphs are constructed to explicitly model cross-view dependencies, and a semantic-level attention mechanism adaptively fuses shared and complementary cues. This design mitigates spurious correlations and enhances interpretability. Experiments on the SingFake and WildSVDD datasets show that our method achieves average EERs of 6.87% and 9.62%, respectively, outperforming state-of-the-art baselines. These results validate that incorporating background and pitch information through graph-based logical fusion effectively enhances robustness and generalization in music deepfake detection.

Read on IEEE Xplore