Fusion music combines diverse rhythms, melodies, and stylistic elements to create a complex and creative musical form. However, traditional classification methods often focus on improving overall classification accuracy, neglecting the unique nature of fusion music genres, which results in lower accuracy. To accurately classify fusion music, we must deeply consider the interactions between different features and perform detailed feature processing. In this paper, we propose an interlaced cross-attention feature fusion network, called ICFF-Net. The network aims to achieve optimal decision results by leveraging the advantages of multiple features and adjusting the weights assigned to different genres. Specifically, we extract features from the spectrograms (2D) of the original audio and the accompaniment audio (1D). The temporal and frequency domain features of the 1D-features are concatenated with the 2D-features as two sources. We then construct an interlaced cross-attention module with a dual-layer structure (IC-Attention) to fuse these features. IC-Attention employs cross-attention with two embeddings from different sources, comprehensively considering the impact of both 2D and 1D features on the classification results, achieving comprehensive feature interaction and weight allocation. Additionally, we constructed a FusionMusic dataset to analyze classification bias’s effect on fusion genres. Our method demonstrates superior performance in addressing classification bias.