CoSCF: Gradient-Aware Surrogate Optimization for Task-Oriented Collaborative Sensing

Abstract

Multi-view WiFi sensing enhances robustness over single-link sensing, but transmitting all intermediate features incurs prohibitive communication overhead. Existing compression methods lack mechanisms to quantify per-view task relevance, making uniform compression inefficient for task-oriented sensing. Therefore, we formulate multi-view collaborative sensing as a bi-level problem and propose CoSCF, which optimizes a single-level surrogate of it rather than solving it exactly. During training, a two-pass gradient feedback mechanism backpropagates the task objective to measure the discriminative contribution of each feature channel and guides edge devices to perform differentiable, adaptive bit allocation and channel sparsification. Experimental results demonstrate that under strict bandwidth constraints, CoSCF achieves outstanding task performance and substantially outperforms existing methods based on data reconstruction and feature compression.