Integrated Sensing and Communication (ISAC) is acknowledged as a significant candidate for sixth-generation mobile communications (6G). A key topic in ISAC is communication-assisted sensing, in which enhanced accuracy and comprehensiveness of sensing are achieved through the optimization of collaboration and communication among sensing nodes. In this framework, federated learning (FL) is extensively utilized to enhance the collaborative capabilities of the system. However, due to constraints imposed by communication resources and bandwidth, training at FL nodes may be interrupted, leading to performance limitations. To alleviate the parameter transmission burden between nodes and the central server, an efficient model compression method based on Tucker decomposition is proposed in this paper. Previous compression methods based on principal component analysis (PCA) were limited to two dimensions, while Tucker decomposition can decompose four-dimensional CNN model parameters into low-rank core tensors and factor matrices. Experimental validation on FL tasks indicates that, compared to PCA-based compression method, the proposed approach not only substantially diminishes communication overhead but also preserves high accuracy. This suggests that Tucker decomposition seems a promising scheme for efficient FL.