Efficient multiscale attention feature infusion for enhancing MAC protocol identification in underwater acoustic networks

The identification of MAC protocols in Underwater Acoustic Networks (UANs) presents notable challenges due to the complexities inherent in the ocean environment. Traditional approaches, which depend on statistical features of received power, often fail to capture latent, nonlinear, deep semantic features. To overcome this limitation, we introduce an Efficient Multiscale Attention (EMA) feature infusion framework. Our approach utilizes a Convolutional Neural Network (CNN) equipped with the EMA block to extract implicit time–frequency domain features. These are then infused into a handcrafted explicit feature set, expanding the feature space for classification using conventional classifiers. This lightweight solution offers significant advantages in edge computing for ocean sensing, compared to more complex models that incorporate handcrafted features directly into neural networks for training. The proposed EMA block enriches time–frequency receptive fields and refines channel attention through local inter-channel interactions, reducing parameter overhead. Experimental results demonstrate the effectiveness and robustness of the proposed method.