Through the coupling and coexistence of sensing and communication functions, integrated sensing and communications (ISAC) has proven to be a viable approach to both eliminate the hardware overhead problem and deliver a richer and more comprehensive sensing experience. Among them, communication-assisted sensing is one of the current research hotspots. It can transmit and receive additional information using communication systems to assist the sensing system in a more accurate and comprehensive understanding of the environment. Therefore, in recent years, federated learning (FL), which can use distributed nodes for sensing, has been widely discussed. However, the convolutional neural network (CNN) using FL will cause huge communication over-head due to a large number of parameters and even affect the sensing performance. To address the problems listed above, an innovative federated transfer learning (FTL) utilizing depthwise separable convolutional network (DSC) is presented in this paper. More specifically, DSC is introduced to reduce the number of parameters in the model and thus reduce communication overhead once a pre-trained model is initially obtained by data augmentation to extract features to expedite learning. Finally, we simulate and analyze the proposed scheme. The simulation results demonstrate that, in terms of average accuracy and communication overhead, our model performs better than the baseline.