TL;DR
This paper introduces a novel framework for mobile crowdsensing (MCS) that combines blockchain and federated edge learning to address issues of resource constraints for requesters and privacy concerns for participants. By leveraging these technologies, the system enables efficient, privacy-preserving data analysis and incentivizes participation.
What Problem Does This Paper Address?
Mobile crowdsensing (MCS) relies on a large number of mobile devices to collect data for various tasks, offering flexibility and cost savings. However, traditional MCS faces significant hurdles: the immense communication overhead for raw data transmission and the substantial storage and computational demands placed on requesters. These limitations prevent resource-constrained users from utilizing MCS, hindering its widespread adoption. Furthermore, privacy concerns arise from the involvement of both requesters and workers in the learning process, and the potential for malicious actors (edge servers or requesters) to disrupt the system.
Methods
The proposed solution integrates blockchain technology with federated edge learning (FL) into a four-entity framework: requesters, a consortium blockchain, edge servers, and mobile devices (workers). The system operates through four key procedures:
- Task Publication: Requesters submit MCS learning tasks to the blockchain via a smart contract, which then recruits workers.
- Data Sensing and Submission: Mobile devices collect data and send it to nearby edge servers. This localizes data processing, enhancing privacy by preventing raw data from reaching a central platform. Edge servers also provide immediate rewards to devices.
- Learning to Return Final Results: Edge servers collaboratively train a machine learning model using federated learning on their local datasets. This process is secured by a consortium blockchain, which acts as a decentralized coordinator, replacing a central aggregator and mitigating single points of failure.
- Payment Settlement and Allocation: Once the learning task is complete, the requester is charged, and incentives are distributed to all participants (mobile devices, edge servers, and blockchain nodes). A zero-determinant (ZD) game theory-based payment enforcement scheme is employed to ensure requesters pay the full amount.
Key novelties include:
- Mechanism Design for Data Submission: A game rule is devised to incentivize edge servers to protect the privacy of submitted data, preventing leakage.
- Consortium Blockchain-based FL: This secures the distributed learning process, with blockchain nodes coordinating and validating updates.
- ZD Game Theory for Payment Enforcement: This strategy ensures requesters fulfill their payment obligations, safeguarding participant incentives.
Datasets
The paper mentions using a benchmark named LEAF [52] for simulating the federated learning process, specifically with the FEMNIST dataset. This dataset is used to train a 2-layer convolutional neural network (CNN) classifier.
Key Insights and Ideas
- Decentralization is Key: By using a consortium blockchain, the system avoids the single point of failure associated with centralized aggregators in traditional FL and MCS.
- Privacy at the Edge: Localizing data processing to edge servers, rather than sending raw data to a central entity, significantly enhances the privacy of mobile device users.
- Incentive Compatibility is Crucial: The mechanism design for data submission and the ZD game theory for payment settlement are critical for ensuring all participants are motivated and fairly compensated, making the system sustainable.
- Trade-offs in FL: The experiments highlight that while more data generally leads to better accuracy, increasing the number of participants (edge servers) can slow down convergence due to smaller local datasets per participant.
Key Results
- Data Submission Mechanism: The proposed mechanism design is proven to be incentive-compatible, ensuring edge servers report their true data leakage degree to maximize their utility.
- Federated Learning Performance: Experiments show that higher dataset fractions improve testing accuracy, while more edge servers lead to slower convergence.
- Blockchain Consensus Efficiency: The time cost for blockchain consensus increases linearly with the number of edge servers and near-exponentially with the number of blockchain nodes.
- Payment Settlement: The ZD-based payment settlement strategy effectively drives the requestor towards full cooperation (paying the full incentive), outperforming traditional strategies like Tit-for-Tat and Win-Stay-Lose-Shift. The system achieves a stable state of mutual cooperation.
Limitations and Future Work
- Scalability of Blockchain: While a consortium blockchain is used, the paper acknowledges that further work is needed to improve its efficiency, scalability, and reliability.
- Security Challenges: The authors plan to investigate security challenges in blockchain-based distributed learning during the model training process.
- Parameter Sensitivity: The paper notes that while various parameter settings were evaluated, some are omitted for brevity. Future work could explore these more extensively.
- Real-world Deployment: The current work is based on simulations. Future research could focus on real-world deployment and testing.
Summary updated