Internal multiple suppression of seismic data has been a research hotspot and difficulty in the field of oil and gas exploration. The strong reflection interface in the subsurface will form internal multiples with strong energy, which seriously affect the identification of primaries. It also can reduce the authenticity and reliability of seismic imaging. The deep learning-based multiple suppression method can form more abstract high-level features by combining the low-level features to better discover the effective features of the data, and the multiple separation accuracy is high. In this paper, the attention mechanism is introduced for the problem of high training cost of traditional convolutional neural network, and an internal multiple suppression method based on the attention mechanism is proposed to reduce the training cost of neural network model. The data test shows that the method is not affected by the limitations of the traditional internal multiple suppression method and can avoid the regularization of seismic data, thus reducing the computational burdens and improving the computational efficiency, which has important theoretical and industrial application value.