Cellular-Assisted, Deep Learning Based COVID-19 Contact Tracing
Cellular-Assisted, Deep Learning Based COVID-19 Contact Tracing
复制标题
DOI：
10.1145/3550332
复制
发表时间：
2022-09
影响因子：
--
通讯作者：
Fan Yi;Yaxiong Xie;Kyle Jamieson
中科院分区：
文献类型：
--
作者：
Fan Yi;Yaxiong Xie;Kyle Jamieson
冠状病毒病（Covid-19）大流行造成了全球的社会和经济危机。接触示踪是一种有效的有效方法，它包含了COVID-19的传播。在本文中，我们提出了基于蜂窝网络通道状态信息（CSI）测量值的基于细胞辅助深度学习的Covid-19触点跟踪系统。 CAPER利用基于Deep神经网络的特征提取器将细胞CSI映射到神经网络特征空间，在该空间中，点之间的欧几里得距离与设备的接近性密切相关。通过这样做，我们通过确保Caper永远不会向其服务器或其他客户端传播一个客户的CSI数据来维护用户隐私。我们使用软件定义的无线电平台实现了刺山柑原型，并在包括室内和室外场景，拥挤和稀疏环境在内的各种现实情况下评估其性能，并具有不同的数据流量模式和常见的蜂窝配置。 Microbenchs表明，我们的神经网络模型在OnePlus 8智能手机上以12.1微秒的速度运行。端到端的结果表明，CAPER的总体准确性为93.39％，在确定两个设备是否在6英尺范围内，只能错过1.21％的紧密联系，从而超过了基于BLE的方法的准确性14.96％。 Caper对环境动力学也很强，在运行十天后的准确度为92.35％。
The Coronavirus disease (COVID-19) pandemic has caused social and economic crisis to the globe. Contact tracing is a proven effective way of containing the spread of COVID-19. In this paper, we propose CAPER, a Cellular-Assisted deeP lEaRning based COVID-19 contact tracing system based on cellular network channel state information (CSI) measurements. CAPER leverages a deep neural network based feature extractor to map cellular CSI to a neural network feature space, within which the Euclidean distance between points strongly correlates with the proximity of devices. By doing so, we maintain user privacy by ensuring that CAPER never propagates one client's CSI data to its server or to other clients. We implement a CAPER prototype using a software defined radio platform, and evaluate its performance in a variety of real-world situations including indoor and outdoor scenarios, crowded and sparse environments, and with differing data traffic patterns and cellular configurations in common use. Microbenchmarks show that our neural network model runs in 12.1 microseconds on the OnePlus 8 smartphone. End-to-end results demonstrate that CAPER achieves an overall accuracy of 93.39%, outperforming the accuracy of BLE based approach by 14.96%, in determining whether two devices are within six feet or not, and only misses 1.21% of close contacts. CAPER is also robust to environment dynamics, maintaining an accuracy of 92.35% after running for ten days.