Abstract: With the rapid development of mobile internet, the number of mobile devices has sharply increased, leading to a shortage of spectrum resources. Dynamic spectrum allocation is an effective way to alleviate the shortage of spectrum resources. Dynamic spectrum allocation relies on spectrum sensing, which detects unoccupied frequency bands. Traditional spectrum sensing methods only consider the scenario of a single sensing node, which can only monitor a limited geographical scope. In order to monitor a large geographical range, it is necessary to consider the joint sensing architecture, namely, distributed spectrum sensing architecture. Due to the varying hardware accuracy and sensing environment, there exists heterogeneity among devices in a distributed system. To address this issue, this paper proposes a new distributed spectrum sensing architecture. The proposed architecture consists of several sensing nodes, each equipped with a convolutional neural network (CNN) to identify whether the spectrum is occupied. The proposed distributed architecture requires shallow layers of different nodes to share weight parameters, while the parameters of the deep layers remain independent across nodes. The motivation stems from the fact that feature extraction in shallow layers is less affected by signal-to-noise ratio (SNR), thus sharing the weights of shallow layers among different nodes can improve sample efficiency. The parameters of the deep layers are more significantly affected by the SNR. To enhance the robustness of the perceptual system to the local SNR, each node should train its deep layers using only its own training samples. The proposed method can significantly improve the detection accuracy of heterogeneous sensing networks.