User Data Privacy Protection Based on Machine Learning and Encryption Algorithms

ABSTRACT The framework adopts a double trapdoor function to encrypt the gradient (first trapdoor) and decrypt it in stages (second trapdoor) to ensure that the original data is still irreversible when a single trapdoor is leaked. By locally training the encryption gradient parameters in federated learning clients, combined with dynamic group aggregation (grouping by similarity) and a global update mechanism, the model robustness is optimized under data distribution deviation and malicious attack scenarios while protecting privacy. The experiment outcomes denote that in the simulation environment, the average accuracy of the model after convergence is 93.24%. Combined with grouping optimization, the average accuracy of the model reaches 96.88%. The average safe calculation time of the model is 11.78 min, which has been optimized to 9.80 min. When the proportion of malicious clients is 10% and 30%, the average accuracy of the optimized model is 93.26% and 83.17%, respectively. In addition, in actual performance testing, the actual average accuracy of the model is 85.92%, and after optimization, the average accuracy increases to 90.14%. The actual average safe calculation time of the model is 16.83 min, and after optimization, its average time is 13.92 min. Research can improve the efficiency of user data privacy protection and enable the application of the big data industry in a wider range of fields.