Abstract
Mobile Crowdsensing (MCS) is a powerful paradigm for large-scale environmental monitoring, yet traditional cloud-centric models face significant hurdles regarding scalability and user privacy. This paper presents EdgeSense, a decentralized crowdsensing framework that utilizes edge computing and peer-to-peer (P2P) networking. By implementing decentralized matrix factorization via parallelized stochastic gradient descent (PSGD), EdgeSense reconstructs spatial-temporal maps from sparse, locally-held data. Evaluations on real-world datasets show that EdgeSense provides high-accuracy monitoring while mathematically guaranteeing that raw sensor and location data never leave the participant's control.
Introduction
Urban monitoring of phenomena like air quality and traffic flow increasingly relies on Mobile Crowdsensing (MCS). However, the prevailing "cloud-centric" architecture requires a central server to aggregate massive amounts of data, leading to processing bottlenecks and severe privacy risks. In these models, service providers often require precise time and location data, allowing for the potential tracking of user trajectories.
EdgeSense addresses these issues by shifting the paradigm from centralized aggregation to decentralized edge-mediated processing. The framework aims to provide the benefits of "big data" environmental monitoring without the associated privacy costs.
Methodology
The EdgeSense framework is built upon a decentralized architecture that removes the need for a central data hub.
System Architecture
The system consists of mobile participants, trusted edge nodes, and a P2P network. Each participant pairs their device with a trusted edge node (such as a home computer), which acts as a proxy to strip identifying information like IP addresses before communicating with the rest of the network.
Decentralized Matrix Factorization
EdgeSense models the sensing area and time as a large matrix where most cells are empty. To fill these gaps, it employs Non-negative Matrix Factorization (NMF). Instead of sending raw data to the cloud, participants perform local updates on "latent factors" (mathematical representations of data trends) using Parallelized Stochastic Gradient Descent (PSGD). These factors are then shared across a Gossip-based P2P network.
Three-Phase Workflow
- Initialization: The organizer establishes the P2P network and parameters.
- Distributed Optimization: Participants update matrix factors locally using their private sensor data and pass these updates to peers.
- Recovery: The organizer receives the final averaged factors and reconstructs the full environmental map.
Experimental Evaluation
The researchers evaluated EdgeSense using two real-world datasets to simulate urban sensing: TEMP (temperature readings from 57 subareas) and PM25 (air quality data from 36 stations).
Key Results
- High Accuracy: EdgeSense achieved a low approximation error—roughly 0.2°C for city-wide temperature maps and 10 units for PM2.5 indices (on a scale of 1 to 500).
- Scalability: As the number of participants increased from 10 to 40, the error rate consistently dropped, demonstrating that the system thrives on crowd participation.
- Performance Parity: The algorithm performed comparably to Spatio-Temporal Compressive Sensing (STCS), the leading centralized baseline, while significantly outperforming other standard methods like Robust PCA (RPCA) and Truncated SVD.
Conclusion
EdgeSense demonstrates that decentralized crowdsensing is a viable and competitive alternative to cloud-centric models. By sharing latent factors rather than raw sensor or location data, the system provides a mathematical guarantee of privacy. While the P2P communication overhead and the requirement for trusted edge servers present areas for future optimization, the framework establishes a robust foundation for privacy-preserving, large-scale spatial-temporal sensing.
Summary updated