1. Improving Location Reliability in Crowd Sensed Data with Minimal Efforts
Authors: Manoop Talasila, Reza Curtmola, and Cristian Borcea
Presenter: Hillol Debnath
Department of Computer Science
New Jersey Institute of Technology
•This work is focused on Mobile crowd sensing
•Static sensors were used Previously to sense data from the physical world (e.g. traffic monitoring, environment/weather monitoring etc.). Which was much more expensive and not scalable.
- Everyone is using smartphones these days. So, people centric sensing can be used for large scale sensing of the physical world by leveraging smartphone sensors such as camera, microphone, GPS, accelerometer, etc. This type of sensing is less costly and more scalable than the static sensing. - However, one major problem with these sensed data is - how to validate them whether they are fake or real.
•Why we need to validate the data? … Complex problem….And we have alleviated the validation problem. (professor’s mail)

2. Background – Crowd Sensing
Crowd Sensing: Scalable and cost effective coverage of large regions
Crowd
Sensing
Crowd
Sensing
Source: IDC
Sensed Data
Camera
Accelerometer
GPS
Bluetooth
Microphone
Let me start with an overview of crowd-sensing. (talk about the stat)
Talk about sensors on smartphones. How we can use sensors for collecting task data
Using these two features we can achieve crowd-sensing
Map – to show that we can collect data from a very wide region from which collecting data manually would be burdensome.
(Regular people would submit sensor data, the central system would collect data. The 3rd party would be road transportation agency, news orgs, govt orgs like municipal corporations, etc)
Motivation
Problem
Design
Evaluation
Conclusion

3. Examples of Mobile Crowd Sensing
Traffic jam alerts
Citizen-journalism
Environment
Traffic data –
Citizen journalism – people take photos (and then news org can bid for their data) individual app – no centralized sys
Environment sensor – (pollution sensors)
Motivation
Problem
Design
Evaluation
Conclusion

4. How Reliable is Crowd Sensed Data?
Submit “false” traffic jam
Submit “fake” pollution
Need methods to validate sensed data!
(from paper)
Fake traffic data to get benefit by diverting all the traffic to another route
Fake photo can be submitted to earn money easily
Environment – even though it’s clean and green. A malicious person (rival party) can report fake pollution data to hurt the org
Submit photo from “fake” location
Motivation
Problem
Design
Evaluation
Conclusion

5. Challenges
Validating every sensed data point of each participant is difficult and not scalable
This research: mitigate the problem by validating the location associated with the sensed data points to improve reliability
We cannot manually validate huge amount data. Not scalable, and data with high sampling rate (acc) cannot be validated manually
We alleviate the problem by validating the location associated with sensed data points. This improves the reliability of sensed data and provides guarantees to clients that collected data is valid.
However, Validating locations without support from the wireless carriers is difficult.
Location validation without support from the wireless carriers is difficult
Motivation
Problem
Design
Evaluation
Conclusion

6. Contributions
Proposed ILR scheme for location validation of crowd sensed data
Developed McSense system & Android application for crowd sensing
Ran two-months user study of McSense and evaluated ILR scheme on the collected data
- To solve this problem we propose ILR scheme.
We developed McSense – a crowd sensing system and it’s sensing application in android
(note: how to increase reliability, how to schedule tasks efficiently, how increase the penetration rate by motivating people)
We evaluated the ILR scheme on the real time data collected from the user study during a two months interval
We also performed simulations to quantify ILR performance at large scale
Evaluated ILR performance at larger scale through simulations
Motivation
Problem
Design
Evaluation
Conclusion

7. ILR: Improve Location Reliability
ILR starts from a small set of validated photos
Implicitly, their location is validated
User co-location is used to extend the set of validated sensed data points starting from validated photos
Co-location: users have Bluetooth ON and perform periodic scans to detect neighbors
Detect False Claim
Transitive Trust
Photo Validation
ILR has 3 different phases.
Photo selection phase – we start with a small set of validated photos. How we select those photo is discussed in later slides
Extending the trust – we call it transitive trust. Using BT discoveries, we detect neighbors from user co-location
After that we execute validation phase to detect fake location claims A B C
Motivation
Problem
Design
Evaluation
Conclusion

8. Photo Selection & Validation
Using collected data from Bluetooth scans, construct connected graph of co-located data points for a given location and a time interval
From these graphs, select the photo data points that have node degree greater than a threshold
Reduce the cost of photo validation by selecting only the most connected photo data points
Selected photos are validated by
Humans (Crowdsourced to Amazon MTurk)
Computer vision techniques
Validators: Locations of the validated photos which are considered reliable
Don’t talk much about vision techniques
Motivation
Problem
Design
Evaluation
Conclusion

9. Extend Trust Transitive Trust Trusted Trusted Trusted Unknown
Example of how ILR works
One person doing the accelerometer sensing does a BT scan periodically
A second person performs a photo task in te same location and time of the 1st person.
After validating the photo task of 2nd person manually, we consider his location as trusted
Because they see each other in their BT scans, trust can be extended from the person performing the photo task to the person performing the acc sensing
Transitive trust:
Mark co-located data points as trusted
Extend the trust from Validators to other co-located data points found in Bluetooth scans
Trust is extended until all co-located data points are trusted or no other data point is found
Alternately, ILR can set a TTL on extended trust
Trusted
Trusted
Unknown
Validator
Motivation
Problem
Design
Evaluation
Conclusion

10. Detect False Claims Chicago NJIT Student Center Trusted Proven FakeB C
Chicago
NJIT Student Center M1 P3 P2 P3 P4 P1 P2 P4
Proven Fake
Trusted
Claims Location:
“NJIT student 10AM” P2 P3 P1 P1
ILR also applies a basic spatio-temporal correlation check
Trusted
Trusted P5 P1 P2 P4
Trusted
Unknown
 Students found around 10AM
Motivation
Problem
Design
Evaluation
Conclusion

11. Prototype Implementation
McSense: a platform for mobile crowd sensing
McSense application is implemented in Android and deployed to Google Play
4 Tabs provided in the mobile application:
For our research, we developed McSense system
Back end server (glassfish, java, MVC…)
The android application developed and deployed in Google Play. (v 2.2 +)
Explain each tab and relate that to the life cycle of a task
If a task is posted by the admin, it would be shown in the available task tab
Completed: User completed tasks are listed
Available: Available tasks on server are listed daily until they expire
Earnings: User’s earnings for all successfully completed tasks
Accepted: User accepted tasks are listed until they expire
Motivation
Problem
Design
Evaluation
Conclusion

12. Tasks Developed for McSense
Manual Photo Sensing Task
Automated Sensing Task using Accelerometer and GPS Sensors
Automated Sensing Task using Bluetooth radio
Automated Sensing Task for Application/Network/Battery Usage
Automated Sensing Task for WiFi Access Points Recording
Motivation
Problem
Design
Evaluation
Conclusion

13. User Study Demographics:
Deployed participatory sensing tasks in real time to campus students and other participants
Participants downloaded and installed the application on their Android phones
2 months user study at NJIT campus to collect crowd sensed data
Users received monetary compensation based on the number/type of completed tasks
Demographics:
Total participants: 58
Motivation
Problem
Design
Evaluation
Conclusion

14. Real-world Evaluation
Total Photos: 1784
Manually Validated: 204
Extend Trust: 148
25% Data points involving false claims are detected
40% Cheating participants are detected
40% is out of total fake participants (not out if all 58 participants.)
How will ILR perform at larger scale if users are required to have Bluetooth on?
The results are promising given the user study conditions:
Only users collecting co-location data had Bluetooth ON
Users didn’t meet often due to campus size & limited number of users in the study
Motivation
Problem
Design
Evaluation
Conclusion

15. Simulation Results
Say on the first graph what density you used (density = 5)
Say on the second one what’s the percentage false location claims from the total number of tasks (10% fake claims)
(This shows that our system works perfectly at large scale. ) We proved that sys would be befitted by minimizing the number of photo tasks required for manual validation.
even with small number of manually validated photos, we can efficiently detect false claims
We also prove that ILR works at low densities
Significant percentage of false location claims detected with few validations
ILR works well at varying node densities
Motivation
Problem
Design
Evaluation
Conclusion

16. Summary ILR Scheme Prototype Simulation
Improves location reliability of mobile crowd sensed data
Detects false location claims
Prototype
Developed McSense platform for crowd sensing
Evaluated ILR on real-world data
Simulation
ILR works well at various node densities
Low % of photo tasks needed to bootstrap ILR
Motivation
Problem
Design
Evaluation
Conclusion

17. Thank You Acknowledgment: This work was supported by NSF Grant CNS and CNS

18. Related Work
MetroSense, Participatory Sensing, Urbanets, Medusa framework
Trusted Platform Module (TPM)
YouProve
LINK protocol
Motivation
Problem
Design
Evaluation
Conclusion