Recent advances in deep generative models have enabled sophisticated music synthesis, capable of manipulating musical elements in isolation or holistically, thereby posing unprecedented challenges for authenticity verification. Existing audio deepfake detection methods predominantly target the speech domain, with limited research dedicated to music deepfake detection and lacking multi-modal capabilities to address diverse music synthesis techniques. This paper proposes MOSS, a novel framework for music forgery detection that adopts a multi-modal design philosophy by extracting and fusing complementary representations from distinct musical components to achieve comprehensive perception. Specifically, MOSS employs three parallel branches: MERT for music semantic understanding, XLS-R for acoustic feature modeling, and RoBERTa for lyric semantic analysis. The framework implements a sophisticated cross-level fusion strategy that integrates multi-layer representations from each modality through circular integration mechanisms, followed by hierarchical Mamba blocks that capture both intra-modal temporal dependencies and cross-modal correlations. Experiments on the SONICS dataset demonstrate that MOSS achieves state-of-the-art performance in music forgery detection.