Distributed Machine Learning Privacy Protection Algorithm for Tax Big Data Analysis

With the acceleration of informatization and digitization, the tax system has generated massive amounts of data with diverse types and large scales. However, the data sharing across regions and institutions faces challenges on privacy protection and compliance. Therefore, a distributed machine learning privacy protection algorithm for tax big data is proposed, and a multi-layer secure transmission mechanism combining differential privacy, homomorphic encryption, and secure multi-party computation is designed. In the experiment, real invoices and tax declaration data from provincial tax bureaus, as well as simulated data generated based on these data, are selected to compare various existing methods. The results showed that the accuracy in classification tasks reached 0.87, which was 2.35%–8.75% higher than that of traditional distributed methods. In the regression task, the mean square error and mean absolute error were reduced by 10%–40% and 22.03%, respectively. Compared to homomorphic encryption methods, the designed method reduced communication overhead by 59.67% and achieved a fault tolerance of 96.38% under the 10% node dropout rate. In addition, the accuracy decrease in poisoning attack scenarios was only 25.29%, which was superior to other methods. This algorithm can achieve high predictive performance and robustness while ensuring privacy and compliance, providing effective technical support for intelligent tax governance.