Abstract: To address the challenge of non-cooperative underwater acoustic communication signal recognition in time-varying channel environments, an automatic modulation recognition method based on an attention-based residual network is proposed in this paper. The method employs a residual structure as the backbone, designing a lightweight network model suitable for underwater platforms, which alleviates the vanishing gradient problem through cross-layer connections. Additionally, an attention mechanism is introduced to enhance the extraction of modulation-sensitive features, thereby improving the modulation recognition capability of the model. Experimental results demonstrate that the model achieves a recognition accuracy of 94.3% and 93.9% on simulated and real underwater acoustic communication signal dataset, with an average improvement of 3.1% compared to baseline models. On specific datasets, the recognition accuracy reaches an average of 97.8%. The model has only 0.26M parameters and achieves a processing speed of 0.61 ms per frame on the training platform. Furthermore, the model supports transfer learning, achieving a recognition accuracy of 92.7% on newly added datasets.