Secure and Intelligent Mobile Crowd Sensing

Transcription

1 Secure and Intelligent Mobile Crowd Sensing Chi (Harold) Liu Professor and Vice Dean School of Computer Science Beijing Institute of Technology, China June 19, 2018 Marist College

2 Agenda Introduction QoI aware, energy efficient participant selection Human-intervention aware participant selection AI for Crowd Sensing Centralized solution Distributed solutions Blockchain-enabled Distributed solutions 2

3 Concept What s mobile crowd sensing/participatory sensing? People use various sensors of smart devices freely to collect different sensing data in different areas, this campaign is called as Participatory Sensing. Three elements of a participatory sensing system: Participant: who participates in collecting sensing data. Task publisher: who needs sensing data. Platform: recruit participants, process participants sensing data, send results to task publisher, and provide reward. 3

4 Applications Emergent Event Detection (e.g., crowd) Smart Transportation Systems (e.g., accident recovery) 4

5 Figure: Quality comparison between the crowdsourced panorama (a) and the ground truth panorama (b). Chen et al. designed a smartphone-empowered crowdsourcing system named IndoorCrowd2D for indoor scene reconstruction. S. Chen, M. Li, K. Ren, X. Fu, C. Qia, "Rise of the Indoor Crowd: Reconstruction of Building Interior View via Mobile Crowdsourcing," ACM SenSys

6 Figure: ground truth, outdoor experiment site Figure: RSS map construction of proposed method, comparing other method. Xiang et al. presented a scheme named CARM that exploited crowd-sensing to construct outdoor RSS maps using smartphone measurements. C. Xiang, P. Yang, C. Tian, L. Zhang, H. Lin, F. Xiao, M. Zhang, Y. Liu, "CARM: Crowd- Sensing Accurate Outdoor RSS Maps with Error-Prone Smartphone Measurements," in IEEE Transactions on Mobile Computing, vol. 15, no. 11, pp , Nov

7 Figure: MobiGroup system architecture Guo et al. presented a group-aware mobile crowd sensing system called MobiGroup, which supported group activity preparation, including a heuristic rule-based mechanism for advertising public activity and a context-based method for private group formation. B. Guo, Z. Yu, L. Chen, X. Zhou and X. Ma, "MobiGroup: Enabling Lifecycle Support to Social Activity Organization and Suggestion With Mobile Crowd Sensing," in IEEE Transactions on 7 Human-Machine Systems, vol. 46, no. 3, pp , June 2016.

8 Overall Scenario Multi-task system in a 2D/3D sensing region Multi-dimensional quality-of-information (QoI) requirements Remaining energy level and the willingness for participation 8

9 QoI-Aware Energy-Efficient Participant Selection 9

10 System Model An example about the sensing accuracy: Obviously, different distributions of even same amount of samples can lead to different mean errors 10

11 QoI Satisfaction Metric QoI Satisfaction Metric: amount of required data amount of collected data Then, QoI satisfaction metric can be calculated as 11

12 Energy Consumption Index denote the impact when performing a particular sensing task q on a participant s smart device s energy consumption. The exponent factor > 0 denotes the degree of user disturbance Weight factor k scales the index proportionally Online questionnaire results to fit parameters 12

13 Optimization Problem The goal is to find a subset of participants who can both optimally achieve the QoI requirement of a sensing task, and suffer minimum energy cost during the data collection. We formulate the following multi-objective constrained optimization problem: 13

14 Simulation Setting We use the GeoLife GPS Trajectories of Microsoft Research. We store all trajectories in a MySQL database and find a region (200 by 500 meters) that is of high movement density. The selected region is divided into 160 subregions. The area of every sub-region is 25 square meters (5x5). All 612 trajectories in our database are taken as candidate participants. 14

15 Results Impact of avr. Initial energy reserve on data collection, that 70%-100% energy will increase 48% more data. Impact of energy cost of each task on data collection, that compared with random selection, our proposal increases 25% more data Impact of required amount of data on data collection, that compared with random selection, our proposal increases 20% more data 15

16 Human-Intervention Aware Participant Selection 16

17 New Challenge! The central server needs sensory data with different QoI requirements: Multi-dimensional High degrees What we have covered The willingness for participation of participants depends on various factors: Incurred device energy consumption Impacts on their regular activities We have NOT covered How to specify a generic sampling behavior for all participants due to their diversity and mobility patterns, considering potential human interventions? 17

18 Energy-Aware Sampling Behavior Model A large number of recommended samples means that the participant may collects more sensory data with higher rejection probability. Relationship between the rejection probability, recommend number of samples and the remaining energy of devices can be modeled and confirmed with online questionnaires. results from online questionnaires 18

19 Optimization Problem The objective is to find an optimal crowd of participants to collect sufficient amount of sensory data with required data quality and limited task budget. 19

20 Performance Evaluation Data set: GPS trajectory dataset collected in (Microsoft Research Asia) Geolife project. w/o considering the rejection probability, platform will face high QoI loss: QoI loss in common algorithm: 67.3% QoI loss in proposed algorithm: 6.8% No. of required samples for each participant is not always the more the better, an optimal value can be found by careful analyzing of the participants sampling behaviors. 20

21 AI for Mobile Crowd Sensing 21

22 Reinforcement Learning in a nutshell 22

23 Deep RL 23

24 RL in detail 24

25 DQN in Atari 25

26 A3C in Labyrinth 26

27 DDPG for Continuous Control 27

28 We Start from a Simple Case 1 UAV Simple Case: one UAV is controlled to collect data, while the destination is a charging point. 28

29 DQN Model 29

30 Dataset and Trajectory 30

31 Simulation Results 31

32 Multi-Agent Decentralized Control 32

33 Model 33

34 2 UAV Trajectories 34

35 Multi-Agent DDPG Secure Decentralized Control Endor se( TX) Endor ser Peer s Appl i cat i on/ SDK Respond( Endor sed TX) Del i ver Bl ocks Broadcast ( Endorsed TX) Orderers 35

36 Thank You! Professor Chi (Harold) Liu Vice Dean, School of Computer Science Beijing Institute of Technology, China