SDAR-Net: A Low-Power Solution to EEG-Based Auditory Attention Detection

Auditory Attention Detection (AAD) from EEG signals is an enabling technology for next-generation hearing aids and brain-computer interfaces. However, deploying state-of-The-Art deep learning models on resource-constrained wearable devices is hindered by high computational and energy cost. To overcome this challenge, we propose SDAR-Net, a novel, fully spiking neural network (SNN) designed for highly efficient and accurate AAD. SDAR-Net synergistically combines a dualpathway convolutional front-end for rich feature extraction with stacked transformer-based spiking attention blocks for hierarchical feature refinement. Evaluated on a challenging audiovisual EEG AAD dataset, SDAR-Net achieves state-of-The-Art accuracy of 84.9 percent, outperforming established artificial neural networks, while reducing energy consumption by over 62 percent. Our work showcases the potential of deep SNNs to enable high-performance, low-power neurotechnologies for realworld applications. The source code for SDAR-Net will be made publicly available.