Asymmetrical Attention Network for Multi-Task WiFi-Based Sensing

WiFi-sensing systems that can accomplish multiple tasks simultaneously are attracting significant attention due to their potential for large-scale commercial applications. However, different WiFi sensing scenarios may often rely on various task-specific features, posing a challenge in balancing these different, or asymmetrical, characteristics across tasks. In this paper, we propose a system that aims to address the asymmetrical problems in the joint recognition of users' locations and activities. First, we define activity recognition as a high-level task and location recognition as a low-level task based on their respective difficulty levels. Then, the proposed system employs cascading attention-based modules to transfer prior knowledge between different tasks. The key insight of the proposed architecture is to mimic skilled learners in similar situations, who often tackle easier problems first to enable them to solve more challenging problems later on. Based on this behavioral strategy, the proposed attention-based modules are designed to generate masks that select specific characteristics from the low-level task to help the high-level task learn respective features more effectively. Finally, extensive experimental results based on two open datasets demonstrate the superiority of our system in accuracy compared to other state-of-the-art methods.