We are a research group based in the School of Computer Science and Engineering, Nanyang Technological University. We focus on the research, design, and evaluation of networked, energy-efficient, and secure sensing systems found in the Internet of Things (IoT) and its AI-empowered generation (AIoT). Our research has two main sub-directions of IoT sensing systems/applications and security/privacy of AIoT sensing. In the first sub-direction, with a strong experimental focus, we study a number of sensing modalities (e.g., powerline radiation, radio frequency, acoustics, image, thermal, and energy), exploit them to construct system functions and applications. In the second sub-direction, we study the security and privacy of AIoT sensing systems that use machine learning to process the sensed data. Due to the immediate application potential, our research has been funded externally by government authorities and companies in the ICT, energy, and manufacturing sectors.
Main contributors from the group: Linshan Jiang (topic coordinator), Chaojie Gu, Mengyao Zheng (alumnus), Dixing Xu (alumnus)
As explained in our post, a hybrid computing paradigm consisting of edge computing at the front end and cloud computing at the back end will prevail along with the formation of IoT as a global infrastructure. In addition, the deep neural network-based learning and inference will be important for improving the sensing performance of IoT systems. In a number of scenarios, the IoT edge and the cloud back end need to work together to implement AI-empowered sensing, during which privacy-sensitive data generated at the edge may be exchanged between the edge and the cloud. The group has ongoing research on designing and evaluating privacy-preserving learning and inference approaches in AIoT systems. Our challenge paper (PDF) provides a taxonomy of the existing privacy-preserving learning and inference approaches.
Main contributors from the group: Qun Song (topic coordinator), Zhenyu Yan
Image modified from this, credit goes to the source. Also refer to this paper for background
By 2025, it is estimated that there will be more than 41.6 billion networked IoT devices. These IoT devices generate 79.4 zettabyes data yearly, which is almost twice of today’s whole Internet (44 zettabytes). Transmitting such massive IoT data to the clouds for centralized processing will face communication bandwidth bottleneck. In addition, the communication networks will introduce uncertain time delays that are undesirable for many applications. To cope with these grand challenges, edge computing moves the computation on the IoT data closer to the data sources. Specifically, the IoT sensors and the last-mile data aggregators, acting as edge nodes, will undertake a significant portion of the computation on the IoT data; only data summaries and commands will be exchanged between the edge nodes and the cloud servers through the Internet when needed. As such, a hybrid computing paradigm consisting of edge computing at the front end and cloud computing at the back end will prevail along with the formation of IoT as a global infrastructure.
GhostStripe is a physical adversarial attack system that uses LEDs to exploit the rolling shutter effect in cameras, generating adversarial color stripes that consistently mislead traffic sign recognition – while remaining invisible to human observers. Published at ACM MobiSys 2024, GhostStripe was also demonstrated at ACM SenSys 2024, where it won the Best Demo Award.
The BubCam system presented in this paper estimates the volume of the air bubbles in the ink bags moving on the conveyor belt of the manufacturing line. It employs one main camera and several wireless cameras, whose operations are controlled by a deep reinforcement learning agent. BubCam achieves 1.34 accuracy improvement and 260x latency reduction, compared with manual inspection.
This paper associates the sensing results of mmWave radar and user-held IMUs to track multiple users moving in the same space and estimate the interpersonal distances.
This paper uses a smartphone’s loudspeaker to emit near-inaudible chirps and its microphone to record the acoustic echoes from the indoor environment for localizing the smartphone. Contrasive learning is applied to learn effective location-discriminative features for constructing a SLAM system.
The PriMask system uses personalized lightweight neural networks to obfuscate the data generated by individual users to protect their privacy against curious cloud inference services. At the same time, the obfuscation does not degrade the original inference service quality.
The Telesonar system employs an active sensing approach to probe the existence of the echo channel at the remote end of a phone call as well as a passive sensing approach to analyze the breath timing. This helps decide whether the call is a robocall.
This is the repo of the FedCFC code.
This is the repo of paper “Physics-Directed Data Augmentation for Deep Sensing Model Transfer in Cyber-Physical Systems”.
This is the repo for the EWSN'22 paper “Sardino Ultra-Fast Dynamic Ensemble for Secure Visual Sensing at Mobile Edge”.
This repository provides code of “Lightweight and Unobtrusive Data Obfuscation at IoT Edge for Remote Inference”.
This dataset includes sensor measurements collected from an air free-cooled data center testbed in 2018 and 2019.
This dataset includes sensor measurement traces collected from an air-cooled data center testbed in 2022 and 2023.
|
|
Conference and workshop
|
|
|
|
Journal
|
|
|
|
Thesis
|
|
|
|
Book chapters
|
|
|
|
Technical report
|
|
|
|
Our group have a good history of success in working with undergraduate students. We have produced many novel research publications driven by undergrads.
Our group has research positions available regularly. We invite motivated researchers to apply.
