A Blockchain-Based Method for Power-Related Data Quality Auditing

Power big data technology is playing an increasingly important role in the regulation and data analysis of smart grids, which face “data silos” and privacy issues. Federated Learning is a distributed machine learning framework and allows model training to be done without compromising user raw data. To protect private information, the model parameters uploaded by the user are usually trained in cipher text. However, the presence of ciphertext data makes it difficult to audit the quality of data uploaded by users. Smart grid federated analysis tasks are vulnerable to attackers launching attacks such as data poisoning and free-riding, which can have a serious impact on the global model trained. To defend against data poisoning and free-riding attacks, there is a need to audit encrypted grid data uploaded by users. In this paper, we propose a blockchain-based power-related data quality audit method that can ensure the correctness of the smart grid federated analysis model. In particular, we design an efficient noise addition mechanism that makes the aggregation model parameter noise sum to zero, which can protect user data privacy while ensuring the usability of the aggregation model. In addition, we propose a grouped aggregated data quality audit algorithm that can quickly locate users who upload malicious data. We conducted experimental evaluations on both the Real Power dataset and the MNIST dataset, the results showed that our approach is effective against data poisoning and free-riding attacks.