The Network is The Sensor

点击翻译按钮获取中文摘要

The title "The Network is The Sensor" paraphrases the well-known "The Network is The Computer" that launched a paradigm shift to networked workstations as the basis for affordable computing. We pronounce that networking has now also become an integral part of the sensing infrastructure. Apart from transferring data, networking devices can report observations about the network itself to enrich end-user applications. The observations (which we will call "network sensing") can be collected from various layers of the protocol stack. As an example, network layer information, such as IP addresses, are already used for user location inference and targeted advertising. In other applications, overhearing nearby Bluetooth Low Energy transmissions/beacons is used for indoor localization based on collected physical layer Received Signal Strength Indication (RSSI) information. Recently developed contact-tracing applications exploit precisely this type of information. We will exploit this tremendous opportunity by finding novel ways to fuse network sensing with ordinary sensor data. For example, we have reported on one such novel application by fusing visual information and device addresses, such that each tracked individual's image is annotated with the device address(es) they have in their possession at any point in time. The application showcases not just the fusion of network and visual sensing, but also how learning and prediction algorithms can reap benefits from such data fusion. The proposed research program broadens sensing to include a systematic way of fusing network sensing with ordinary data sensing. We will start with a systematic survey of Internet-of-Things literature mapping it to the Joint Directors Lab (JDL) fusion model. We will then address issues of where network sensing provides the most benefit in the data-processing chain, and how we can support it on devices with limited resources. We accept that the limited resources, as well as other extrinsic factors, will result in varying quality and intermittent network sensing. We will develop new algorithms and techniques to address these issues as a combination of compression and data imputation. We will handle heterogeneity via a virtualization layer such that various network sensing sources are combined (possibly together with ordinary sensors as well) into "virtual sensors". We will combine the virtual sensor concept with metrics that allow us to identify when it is beneficial to combine multiple sensors into one virtual and when not. We will recognize the dynamics, e.g. movement, that could require the re-organization of the virtual sensors. Our overall evaluation strategy will be based on large-scale co-simulations, where existing domain-specific discrete event simulators cooperate via a synchronization of their main event loops. Finally, we will release the resulting (co-)simulation infrastructure as an open source project.