I am working with the members of NTU IoT Research Group to conduct research on networked sensing in various cyber-physical systems that are carried by humans, embedded in infrastructures, and distributed in wild environments. With an experimental basis, our research tries to understand how the physical processes affect the cyber systems (sensor networks and the broader Internet of Things) and exploit certain properties of the physical processes to advance the functional and non-functional aspects of the cyber systems. Our research has been recognized by various best paper awards and runner-ups at prestigious international conferences. With immediate application potential, our research has been funded externally by government authorities and companies in the information technology, energy, and manufacturing sectors. I have been serving on the technical program committees of various international conferences and also the editorial boards of journals that are related to networked sensing.
Research interests
AIoT (AI + IoT)
Sensing for resilient system functions (e.g., timing, synchronization, location, etc)
Names underlined are students, research staff, visiting students who worked directly with me in my group for the publication.
[ACM/IEEE ICCPS'23]Jiale Chen, Duc Van Le, Rui Tan, Daren Ho. BubCam: A Vision System for Automated Quality Inspection at Manufacturing Lines. Best Paper Award. [pdf]
[ACM SenSys'22]Linshan Jiang, Qun Song, Rui Tan, Mo Li. PriMask: Cascadable and Collusion-Resilient Data Masking for Mobile Cloud Inference. Best Paper Candidate. [pdf]
[EWSN'22]Qun Song, Zhenyu Yan, Wenjie Luo, Rui Tan. Sardino: Ultra-Fast Dynamic Ensemble for Secure Visual Sensing at Mobile Edge. [pdf] [code]
[ACM/IEEE IPSN'21]Wenjie Luo, Zhenyu Yan, Qun Song, Rui Tan. PhyAug: Physics-Directed Data Augmentation for Deep Sensing Model Transfer in Cyber-Physical Systems. Best Artifact Award Runner-Up. [pdf] [code and data] [video presentation]
[IEEE TSUSC]Duc Van Le, Yingbo Liu, Rongrong Wang, Rui Tan, Lek Heng Ngoh. Air Free-Cooled Tropical Data Center: Design, Evaluation, and Learned Lessons. [pdf]
[IEEE ICDCS'20]Chaojie Gu, Linshan Jiang, Rui Tan, Mo Li, Jun Huang. Attack-Aware Data Timestamping in Low-Power Synchronization-Free LoRaWAN. [pdf] [video presentation]
[ACM SenSys'19]Qun Song, Zhenyu Yan, Rui Tan. Moving Target Defense for Embedded Deep Visual Sensing against Adversarial Examples. [pdf] [slides]
[ACM MobiCom'19]Zhenyu Yan, Qun Song, Rui Tan, Yang Li, Adams Wai Kin Kong. Towards Touch-to-Access Device Authentication Using Induced Body Electric Potentials. [pdf] [slides]
[ACM UbiComp'18]Qun Song, Chaojie Gu, Rui Tan. Deep Room Recognition Using Inaudible Echos. [pdf] [slides]
[ACM SenSys'17]Zhenyu Yan, Yang Li, Rui Tan, Jun Huang. Application-Layer Clock Synchronization for Wearables Using Skin Electric Potentials Induced by Powerline Radiation. [pdf] [slides]
[ACM/IEEE IPSN'17]Yang Li, Rui Tan*, David Yau. Natural Timestamping Using Powerline Electromagnetic Radiation. *Corresponding author. Best Paper Award. [pdf] [slides]
[IEEE RTSS'16]Sreejaya Viswanathan, Rui Tan*, David Yau. Exploiting Power Grid for Accurate and Secure Clock Synchronization in Industrial IoT. *Corresponding author. [pdf] [slides]
[ACM/IEEE ICCPS'16] Rui Tan, Hoang Hai Nguyen, Eddy. Y. S. Foo, Xinshu Dong, David K. Y. Yau, Zbigniew Kalbarczyk, Ravishankar K. Iyer, Hoay Beng Gooi. Optimal False Data Injection Attack against Automatic Generation Control in Power Grids. [pdf] [slides]
[ACM CCS'13] Rui Tan, Varun Badrinath Krishna, David K. Y. Yau, Zbigniew Kalbarczyk. Impact of Integrity Attacks on Real-Time Pricing in Smart Grids. [pdf] [slides]
[IEEE PerCom'13] Dennis E. Phillips, Rui Tan*; Mohammad-Mahdi Moazzami; Guoliang Xing; Jinzhu Chen; David K. Y. Yau. Supero: A Sensor System for Unsupervised Residential Power Usage Monitoring. *Co-primary authors. Mark Weiser Best Paper Award Finalist. [pdf] [slides] [video]
Method and System for Carrying Out Timing Related Tasks. Inventors (ordered alphabetically): Yang Li, Rui Tan, Sreejaya Viswanathan, David Yau.
US Patent 16464264 (applied in May 2019; granted in Oct 2020).
Holistic Moving Target Defence for Autonomous Driving Perception.
Funded by AI Singapore. PI. S$4M, July 2023 to June 2026.
Towards Zero-Carbon Autonomous Driving AI.
Funded by Singapore's Ministry of Education Tier-1 (Thematic Call). PI. S$150K, Mar 2023 to Mar 2026.
Physics-Guided Generalization of AIoT Sensing.
Funded by Singapore's Ministry of Education Tier-1. PI. S$150K, Mar 2023 to Aug 2025.
Air-Cooled Tropical Data Centre 2.0
Funded by Singapore National Research Foundation (NRF). PI. S$1.62M, Apr 2021 to Mar 2024.
As part of an NRF IAF-PP Programme (Sustainable Tropical Data Centre Testbed).
Tropical Edge Data Centre Testbed.
NTU Seed Grant. PI. S$70K, Mar 2021 to Dec 2022.
Practical Touch-Based Access Control for Indoor IoT Objects.
Funded by Singapore's Ministry of Education Tier-1. PI. S$90K, Nov 2019 to Dec 2021.
Location Sensing using Inaudible Echolocation and Powerline Electromagnetic Radiation.
Funded by Singtel Cognitive and Artificial Intelligence Lab for Enterprises. PI. S$567K, Jul 2019 to Dec 2022.
Feasibility of Object Localization using LoRa Radios.
Funded by The Joint NTU-WeBank Research Centre of Eco-Intelligent Applications (THEIA). PI. S$100K, June 2019 to Dec 2020.
AIoT for Predictive Maintenance.
Funded by HP-NTU Digital Manufacturing Corporate Lab. NTU PI. S$3M, Oct 2018 to Oct 2023.
Tropical Data Centre Proof-of-Concept.
Funded by Info-communications Media Development Authority (IMDA) of Singapore. NTU PI. S$1.39M, Dec 2017 to Dec 2019.
Strategic Capability Building for IoT Research.
NTU CoE seed grant. PI. S$120K, September 2016 to August 2018.
Resilient Cyber Infrastructure for Cyber-Physical Systems.
Funded by Signapore National Research Foundation (NRF) as part of Delta corporate lab in NTU. Co-PI. S$600K, July 2016 to June 2019.
Resilient Cyber-Physical Systems by Advanced Sensing and Computing. NTU Start-up Grant. PI. S$200K, Jan 2016 to Jan 2020. Project yield: 19 conference papers and 20 journal papers.
PopSeCo: Power Plant Security by Advanced Sensing and Computing. Funded by Energy Market Authority (EMA) of Singapore. Co-PI. S$2.72M, Apr 2015 to Oct 2018.
Members
Current members:
Duc Phuc Nguyen (PhD student, 2024/01-)
Yihan Xu (PhD student, 2024/01-; Research Associate, 2023/11-)
Zimo Ma (PhD student, 2024/01-; Research Associate, 2023/08-)
Andreas Kuster (co-supervised PhD student)
Gaole Dai (co-supervised PhD student, 2022-) PhD program working title: Contrastive Learning for Wearable Sensing
Dongfang Guo (PhD student, 2021-; Research Associate, 2019-) PhD program working title: Exploit New Sensing Modalities for Location Sensing in IoT
Jiale Chen (PhD student, 2021-; Research Associate, 2020-) PhD program working title: Efficient Designs of Deep Learning Models for IoT Objects
Rongrong Wang (PhD student, 2020-; Research Associate, 2018-2024) PhD program working title: Toward Energy-Efficient and Smart Built Environments via Advanced Sensing
Siyuan Zhou (PhD student, 2020-; Research Associate, 2019-) PhD program working title: Embedded Deep Visual Sensing in Industrial IoT
Yimin Dai (PhD student, 2023-; MEng, 2021-2022)
MEng thesis: Interpersonal Distance Tracking with mmWave Radar and IMUs.
Yuting Wu (PhD student, 2024/01-; Research Associate, 2021-)
Xiangzhong Luo (Research Fellow, 2024/03-)
Wenbin Tian (Visiting Student from ZJU, 2024/12-)
NTU alumni:
Van Duc Le (Senior Research Fellow, 2023-2024; Research Fellow, 2018-2023)
Huatao Xu (co-supervised PhD student, 2021-2024) PhD thesis: Building Generalizable Deep Learning Solutions for Mobile Sensing.
Ruihang Wang (co-supervised PhD student, 2019-2023) PhD thesis: Physics-Informed Machine Learning for Green Data Center Operations.
Wenjie Luo (PhD student, 2019-2023; joined Huawei in 2023) PhD thesis: Exploiting Sensor and Process Characteristics to Tackle Label Scarcity in AIoT Sensing.
Linshan Jiang (PhD student, 2017-2021; Research Fellow, 2022) PhD thesis: Lightweight Privacy-Preserving Deep Learning and Inference in IoT.
Qun Song (PhD student, 2018-2022; Visiting Student @ NTU, 2017-2018; joined the faculty of Delft University of Technology in 2022) PhD thesis: Improving Security of Autonomous Cyber-Physical Systems against Adversarial Examples.
Zhenyu Yan (PhD student, 2016-2020; Research Fellow, 2020-2021; joined the faculty of CUHK in 2021)
PhD thesis: Exploiting Induced Skin Electric Potential for Body-Area IoT System Functions.
Chaojie Gu (PhD student, 2016-2020; Research Fellow, 2020-2021; joined the faculty of ZJU in 2021)
PhD thesis: Exploiting LoRaWAN for Efficient and Resilient IoT Networks.
Xin Lou (Postdoctoral Researcher; lastest development: Assistant Professor at Singapore Institute of Technology)
Dima Dafer Rabadi (co-supervised PhD student; latest development: Assistant Professor at Pennsylvania State University at Shenango)
Subhash Lakshminarayana (Researcher; latest development: Associate Professor at University of Warwick)
Hoang Hai Nguyen (Research Engineer; latest development: Software Engineer at Amazon AWS)
Yang Li (Postdoctoral Researcher; latest development: Alibaba Group)
Sreejaya Viswanathan (Senior Research Engineer)
Zhan Teng Teo (Research Engineer)
Sheng-Yuan Chiu (Visiting Student)
Recent and current research
1. AIoT (2017-now)
Physics-directed domain adaptation
Run-time domain shifts are common in sensing systems designed with deep learning. The shifts can be caused by sensor characteristic variations. Existing transfer learning techniques require substantial target-domain data and thus incur high post-deployment overhead. We study how to exploit the first principle governing the domain shift to reduce the demand on target-domain data. Specifically, we use the first pinciple fitted with few source/target-domain data pairs to transform the existing source-domain training data into augmented data for updating the deep neural networks. We have applied this PhyAug approach to recover the accuracy losses of DeepSpeech2 caused by microphone characteristic using 5-second data collected from the microphone. This work won the IPSN'21 Best Artifact Award Runner-Up.
Moving target defense against adversarial examples
Deep models are vulnerable to adversarial examples. Many existing countermeasures build their security on the attacker's ignorance of the defense mechanisms and can be subverted once the attackers know the details of the defense. In this research, we apply the strategy of moving target defense to generate multiple fork models at run time from a factory-designed base model, that collaboratively detect and thwart adversarial examples. We also develop efficient implementations of our defense on embedded GPU platforms.
Deep learning-based location fingerprinting
This research uses a smartphone to emit a short inaudible acoustic chirp (only two milliseconds long) and record the ambient's reverberation as a fingerprint of the smartphone's location. We apply deep learning to deal with the challenge of the reverberation's limited information due to its narrow band and short recording time (only 0.1 seconds). We have applied this approach to implement a room-level localization system that can recognize up to 50 rooms with 97.7% accuracy. We have also used it to fingerprint 15 locations in a crowded museum and achieve 89% accuracy.
Lightweight privacy preservation for deep learning and inference
Resource-constrained edge devices are normally incapable of sophisticated privacy preservation mechanisms (e.g., homomorphic encryption). We have studied lightweight privacy preservation approaches, including independent Gaussian random projection for collaborative deep learning and data obfuscation using a shallow neural network for cloud inference using a deep neural network.
2. Sensing for resilient system functions, e.g., timing, synchronization, location, etc (2015-now)
Powerline forensic time and secure clock synchronization
The frequency of the alternating current (ac) voltage of power grid has tiny fluctuations over time. The fluctuations at different locations in a power grid are similar. Based on this, we match the fluctuation traces collected by two voltage sensors to identify the offset between their clocks and synchronize their clocks. This approach achieves 10 microseconds synchronization accuracy in a building and 100 microseconds accuracy for two nodes 10km apart. NTP in LAN can only achieve milliseconds accuracy; PTP (Precision Time Protocol) in LAN can achieve microseconds accuracy, but it requires PTP-enabled switches. Our system also gives security against the packet delay attack that is effective against message-exchange-based clock synchronization protocols. We also extend the above idea to study the accuracy of the forensic time derived from the powerline's magnetic field and achieve 150 milliseconds accuracy. This work won the IPSN'17 Best Paper Award.
Wearable clock synchronization and device authentication in ambient electrostatic field
Powerlines induce an electrostatic field oscillating at the power grid frequency (e.g., 50Hz). This field will cause the redistribution of the charges on the human body viewed as an uncharged equipotential conductor. The redistributed charges in return distort the near-body electric field. We use an ungrounded analog-to-digital converter to measure the transient potential difference between the human body and its near field. As any such two wearables capture the same power grid frequency, we use this common frequency with the principle of NTP and achieve milliseconds synchronization accuracy. Moreover, the transient potential difference signals captured by two wearables at close locations on the same human body are similar. We use this to develop a touch-based device authentication system, in which the user with a token device can authenticate himself/herself to a touchable smart object.
Summary: Exploit power network signals for time acquisition and location sensing
Our research has exploited the ac voltages of a power grid, the magnetic and electrostatic fields induced by the powerlines distributed in civil infrastructures to acquire time and location information. Time acquisition: We can synchronize wall-powered and wireless devices distributed in a geographic area served by the same power system (from building area to city area), with sub-millisecond and sub-second accuracy, respectively. We can also synchronize body-contacted devices worn by users distributed in the geographic area, with milliseconds accuracy. Location sensing: We exploit the powerline-induced magnetic field to perform SLAM in the building area. We also exploit the powerline electrostatic potential received by the human body to perform same-body detection for device authentication.
Secure data timestamping in LoRaWAN
LoRaWAN is promising for the applications of collecting low-rate data. Data samples need timestamps to make sense. However, tightly synchronizing the nodes incurs much overhead to bandwidth-limited LoRaWAN. We propose to perform gateway-side timestamping, which saves bandwidth. However, this gateway-side timestamping is vulnerable to a crafty frame delay attack. We conducted experiments in a campus network, and showed that, by setting up two attack devices (collider and eavesdropper), all LoRaWAN end devices in the area of about 50,000 square meters are affected by the attack. To develop attack awareness, we design a LoRaTS gateway based on cheap radio hardware to detect the carrier frequency offset (CFO) changes caused by the frame delay attack.
3. Thermal and energy management in data centers (2011-2014, 2017-now)
TDC1: World's first trial of air free-cooled data center in tropics (completed)
Air free cooling that utilizes natural outside air to cool the IT equipment in data center has been thought infeasible in Singapore's tropical condition with year-round high temperatures and humidity levels. From 2017 to 2019, we designed, constructed, and experimented with an air free-cooled and deeply sensorized data center testbed consisting of three server rooms hosting 12 server racks with 60kW total power rating. Our results show that Singapore's temperatures of up to 37℃ aren't a concern, but the cleaness and relative humidity of the air supplied to the servers need to be well controlled to maintain the servers' reliability. The energy efficiency metrics (e.g., DCiE) obtained on our testbed would be the upper limit in the local context. The findings of our research are documented in an NTU technical report and communicated to the local data center industry via various invited technical presentations. (Read more details on TDC1.)
TDC2: A commercially ready tropical data center (ongoing)
Based on our TDC1 research results, our TDC2 project sets up an actual data center with 400 servers in an enclosed and conditioned building. The main objectives of TDC2 include: 1) To develop a practicable methodology to determine the optimal setpoints for supply air temperature and relative humidity to achieve the highest energy efficiency; 2) To understand the server reliability under the optimal temperature/humidity setpoints and the implication on the cost-benefit relationship. During the project period, the TDC2 is deeply sensorized for research and meanwhile offers commercial computing services. After the project period, TDC2 will be fully handed over to our data center operator partner for offering continued commercial services. TDC2's ultimate goal is to seed high-temperature data centers with improved energy efficiency in Singapore's tropical condition. (Read CNA's news report on TDC2.)
4. Secure sensing and control in smart grids (2012-now)
False data injection attacks on smart grid controls
Recent high-profile cyber intrusions such as Stuxnet and Dragonfly have alerted us to a general class of integrity attacks called false data injection (FDI). In this research series, we studied the FDI attacks on the frequency control, voltage control (in both alternating current and direct current systems), state estimation, and voltage stability assessment of power systems. In 2015, we attacked the frequency control of an NTU microgrid to deviate the 50Hz grid frequency.
Time delay attacks on smart grid controls
Delaying the network transmissions of control/sensor data or misleading the controller to use outdated sensor data is another class of integrity attacks that impose less requirements on the attacker compared with false data injection attacks. In this research series, we studied the impact of such time delay attacks on frequency control, voltage control, power plant process control, and electricity real-time pricing.
IEEE International Conference on Computer Communications (INFOCOM), 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024
International Conference on Embedded Wireless Systems and Networks (EWSN), 2017, 2018, 2020 (poster/demo session), 2021, 2022, 2023
IEEE International Conference on Sensing, Communication and Networking (SECON), 2018, 2021, 2022, 2023, 2024
Annual International Conference on Distributed Computing in Sensor Systems (DCOSS), 2018
IEEE International Conference on Mobile Ad-hoc and Sensor Systems (MASS), 2019, 2020, 2022
IEEE International Conference on Parallel and Distributed Systems (ICPADS), 2014, 2015, 2017, 2018, 2019
IEEE International Conference on Industrial Internet (ICII), 2018, 2019
IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA), 2017, 2021, 2022
Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of Things (AIChallengeIoT), 2021, 2022
ACM Cyber-Physical System Security Workshop (CPSS), 2015, 2016, 2017, 2022
ACM Workshop on Cyber-Physical Systems Security & Privacy (CPS-SPC), 2015, 2016
Workshop on Benchmarking Cyber-Physical Systems and Internet of Things (CPS-IoTBench), 2021
ACM Workshop on Wireless Network Testbeds, Experimental evaluation and CHaracterization (WiNTECH) with MobiCom, 2021, 2022
Workshop on Mobile and Wireless Sensing for Smart Healthcare with MobiCom, 2022
For university and government
Assistant Dean (Executive & Special Programs), appointed by NTU/CCDS, 2024/05-now
Assistant Dean (Research), appointed by NTU/CoE, 2023/06-2024/04
Area Leader (Green Edge), NRF Foundational Research Capabilities exercise on Green Computing, 05/2023-05/2024
Working Group on Development of New Standard on Energy Efficiency of Data Centre IT Equipment, Green IT Technical Committee, appointed by IMDA, 2023-
Working Group on Development and Operation of Data Centre IT Equipment under Tropical Climate, Green IT Technical Committee, appointed by IMDA, 2021-2023
Data Centre Technical Evaluation Panel member, appointed by IMDA, 04/2021-12/2022
Undergraduate Research Experience on Campus (URECA) Coordinator, appointed by NTU/SCSE, 2020/08-2023/07
Singapore's Tropical Data Centre Guidelines Working Group member, appointed by IMDA, 2018-2019
For K-12
S.T. Yau High School Science Award 丘成桐中学科学奖 (Asia), Assessment Panel on Computer Science, 2021, 2022
(Old) News
[06/2023] Singapore Standard 697:2023, where our research team served as a core working group member, has been announced. IMDA news
[04/2020] Appreciated as Distinguished TPC Member of INFOCOM'20.
[07/2019] Our paper on air free-cooled data centers in tropics is accepted to BuildSys'19. The paper receives (highly) positive comments from all Reviewers!
[07/2019] Our paper on counteracting adversarial example attacks on deep visual sensing is accepted to SenSys'19.
[02/2019] Two papers accepted to IoTDI and ICCPS of CPS-IoT Week 2019!
[02/2019] Two papers on exploiting indoor powerline radiation presented at / accepted by MobiCom'18 and MobiCom'19!
[09/2018] We use a smartphone to emit a 2 milliseconds inaudible chirp and record audio for just 0.1 seconds to recognize a room. The paper will be presented on Ubicomp'18.
[11/2017] Our paper on using LoRaWAN to build control plane for multi-hop wireless networks is accepted to INFOCOM'18.
[07/2017] Our paper on using skin electric potentials to synchronize the clocks of wearables is accepted to SenSys'17!
[04/2017] Our IPSN'17 paper won the Best Paper Award! In this research, we show that the electromagnetic radiation from powerlines contains time information with errors down to 50 milliseconds. PhotoNTU/SCSE newsIllinois news
[01/2017] Appreciated as Distinguished TPC Member of INFOCOM'17.
[08/2016] Our paper on secure clock synchronization is accepted to RTSS'16. Our system achieves 0.1ms sync error for two nodes 10km apart and is provably secure against the packet delay attack!
[01/2016] Our paper on attacks against power grid electricity generation control system is accepted to ICCPS'16. In this work, we really attacked a generator of a microgrid to deviate the 50Hz system frequency.
"Keynote: Toward Efficient Edge AI", Forum on Ubiquitous Computing and Intelligence, World AI Conference, July 06, 2024, Shanghai.
"Physics-Informed Machine Learning for Sensing and Control in Cyber-Physical Systems", Croucher Foundation Advanced Study Institute 2023 - AI for Internet of Things, Jul 19, 2023, The Chinese University of Hong Kong.
"Impacts of Increasing Temperature and Relative Humidity Setpoints in Air-Cooled Data Centers", by Duc Van Le and Rui Tan, on Seminar 34 (The Impact of Hot and Humid and Corrosive Environment on Data Center Equipment: Recent Research Activities on Data Centers), 2023 ASHRAE Annual Conference, Jun 26, 2023.
"Updates of Tropical Data Center 2.0", TC 9.9 Meetings, 2023 ASHRAE Winter Conference, Feb 26, 2023 (presented by Duc Van Le on my behalf)
"Resilient Cyber-Physical Systems by Advanced Sensing and Computing", Department of Computer Science, City University of Hong Kong, June, 2017.
"Wireless Sensor Networks for Volcano Monitoring", Earth Observatory of Singapore (EOS), Oct 18, 2016.
"Worst-Cast Resilience Analysis Enabled by Smart Grid Modeling", lecture on Summer School of The French-Singaporean Network in Renewable Energy (CNERGIE), July, 2016, Porticcio, Corsica, France.
Outreach talks and panel discussions
Panelist at The 2nd China Mobile Southeast Asia Regional Cooperation Conference, Aug 1st, 2024, Bangkok, Thailand.
Run-time domain shifts are common in sensing systems designed with deep learning. The shifts can be caused by sensor characteristic variations. Existing transfer learning techniques require substantial target-domain data and thus incur high post-deployment overhead. We study how to exploit the first principle governing the domain shift to reduce the demand on target-domain data. Specifically, we use the first pinciple fitted with few source/target-domain data pairs to transform the existing source-domain training data into augmented data for updating the deep neural networks. We have applied this PhyAug approach to recover the accuracy losses of DeepSpeech2 caused by microphone characteristic using 5-second data collected from the microphone. This work won the IPSN'21 Best Artifact Award Runner-Up.
This research uses a smartphone to emit a short inaudible acoustic chirp (only two milliseconds long) and record the ambient's reverberation as a fingerprint of the smartphone's location. We apply deep learning to deal with the challenge of the reverberation's limited information due to its narrow band and short recording time (only 0.1 seconds). We have applied this approach to implement a room-level localization system that can recognize up to 50 rooms with 97.7% accuracy. We have also used it to fingerprint 15 locations in a crowded museum and achieve 89% accuracy.
Resource-constrained edge devices are normally incapable of sophisticated privacy preservation mechanisms (e.g., homomorphic encryption). We have studied lightweight privacy preservation approaches, including independent Gaussian random projection for collaborative deep learning and data obfuscation using a shallow neural network for cloud inference using a deep neural network.
The frequency of the alternating current (ac) voltage of power grid has tiny fluctuations over time. The fluctuations at different locations in a power grid are similar. Based on this, we match the fluctuation traces collected by two voltage sensors to identify the offset between their clocks and synchronize their clocks. This approach achieves 10 microseconds synchronization accuracy in a building and 100 microseconds accuracy for two nodes 10km apart. NTP in LAN can only achieve milliseconds accuracy; PTP (Precision Time Protocol) in LAN can achieve microseconds accuracy, but it requires PTP-enabled switches. Our system also gives security against the packet delay attack that is effective against message-exchange-based clock synchronization protocols. We also extend the above idea to study the accuracy of the forensic time derived from the powerline's magnetic field and achieve 150 milliseconds accuracy. This work won the IPSN'17 Best Paper Award.
Powerlines induce an electrostatic field oscillating at the power grid frequency (e.g., 50Hz). This field will cause the redistribution of the charges on the human body viewed as an uncharged equipotential conductor. The redistributed charges in return distort the near-body electric field. We use an ungrounded analog-to-digital converter to measure the transient potential difference between the human body and its near field. As any such two wearables capture the same power grid frequency, we use this common frequency with the principle of NTP and achieve milliseconds synchronization accuracy. Moreover, the transient potential difference signals captured by two wearables at close locations on the same human body are similar. We use this to develop a touch-based device authentication system, in which the user with a token device can authenticate himself/herself to a touchable smart object.
Our research has exploited the ac voltages of a power grid, the magnetic and electrostatic fields induced by the powerlines distributed in civil infrastructures to acquire time and location information. Time acquisition: We can synchronize wall-powered and wireless devices distributed in a geographic area served by the same power system (from building area to city area), with sub-millisecond and sub-second accuracy, respectively. We can also synchronize body-contacted devices worn by users distributed in the geographic area, with milliseconds accuracy. Location sensing: We exploit the powerline-induced magnetic field to perform SLAM in the building area. We also exploit the powerline electrostatic potential received by the human body to perform same-body detection for device authentication.
LoRaWAN is promising for the applications of collecting low-rate data. Data samples need timestamps to make sense. However, tightly synchronizing the nodes incurs much overhead to bandwidth-limited LoRaWAN. We propose to perform gateway-side timestamping, which saves bandwidth. However, this gateway-side timestamping is vulnerable to a crafty frame delay attack. We conducted experiments in a campus network, and showed that, by setting up two attack devices (collider and eavesdropper), all LoRaWAN end devices in the area of about 50,000 square meters are affected by the attack. To develop attack awareness, we design a LoRaTS gateway based on cheap radio hardware to detect the carrier frequency offset (CFO) changes caused by the frame delay attack.
Air free cooling that utilizes natural outside air to cool the IT equipment in data center has been thought infeasible in Singapore's tropical condition with year-round high temperatures and humidity levels. From 2017 to 2019, we designed, constructed, and experimented with an air free-cooled and deeply sensorized data center testbed consisting of three server rooms hosting 12 server racks with 60kW total power rating. Our results show that Singapore's temperatures of up to 37℃ aren't a concern, but the cleaness and relative humidity of the air supplied to the servers need to be well controlled to maintain the servers' reliability. The energy efficiency metrics (e.g., DCiE) obtained on our testbed would be the upper limit in the local context. The findings of our research are documented in an NTU technical report and communicated to the local data center industry via various invited technical presentations. (Read more details on TDC1.)
Based on our TDC1 research results, our TDC2 project sets up an actual data center with 400 servers in an enclosed and conditioned building. The main objectives of TDC2 include: 1) To develop a practicable methodology to determine the optimal setpoints for supply air temperature and relative humidity to achieve the highest energy efficiency; 2) To understand the server reliability under the optimal temperature/humidity setpoints and the implication on the cost-benefit relationship. During the project period, the TDC2 is deeply sensorized for research and meanwhile offers commercial computing services. After the project period, TDC2 will be fully handed over to our data center operator partner for offering continued commercial services. TDC2's ultimate goal is to seed high-temperature data centers with improved energy efficiency in Singapore's tropical condition. (Read CNA's news report on TDC2.)
Recent high-profile cyber intrusions such as Stuxnet and Dragonfly have alerted us to a general class of integrity attacks called false data injection (FDI). In this research series, we studied the FDI attacks on the frequency control, voltage control (in both alternating current and direct current systems), state estimation, and voltage stability assessment of power systems. In 2015, we attacked the frequency control of an NTU microgrid to deviate the 50Hz grid frequency.
Delaying the network transmissions of control/sensor data or misleading the controller to use outdated sensor data is another class of integrity attacks that impose less requirements on the attacker compared with false data injection attacks. In this research series, we studied the impact of such time delay attacks on frequency control, voltage control, power plant process control, and electricity real-time pricing.
