UNITE⁺: Privacy-Aware Data Quality Assessment with Public Verifiability via Federated Learning in Blockchain-Empowered Crowdsourcing

Crowdsourcing, as a novel task execution paradigm, leverages the collective intelligence of workers to complete requester-posted tasks. Quality assessment is a critical phase in crowdsourcing due to the reliability variance of publicly recruited workers. Previous studies on crowdsourcing quality assessment have either overlooked the potential privacy breaches stemming from task data or have been susceptible to biased evaluations from malicious assessors. Moreover, the assessment approach is coarse-grained or unitary. In this research, we propose a privacy-conscious and publicly verifiable data quality assessment scheme called UNITE⁺ in blockchain-empowered crowdsourcing, which is resilient against stronger malicious adversaries. UNITE⁺ employs federated learning (FL) to enable worker-side privacy-aware training of an assessment model, acting as an indicator, together with three other indicators for a comprehensive quality assessment. To safeguard against attackers inferring task data from model gradients in FL, we devise a secure model aggregation protocol based on differential privacy and execute it on blockchain to guarantee trustworthy model aggregation. Such a protocol is also resilient to poisoning attacks from malicious workers. For malicious requesters providing inaccurate assessments, we adopt Pedersen Commitment to generate evidence, recorded on-chain for public scrutiny. Comprehensive security analysis demonstrates that UNITE⁺ achieves (ε, δ)-local differential privacy. Finally, we conduct extensive experiments over Hyperledger Fabric on two real-world datasets, demonstrating satisfactory accuracy, robustness, and on-chain and off-chain performance.