Federated Learning Approaches for Enhanced Chinese Text Classification Across Diverse Domains

This work assesses the advancement achieved in cross-domain Chinese text categorization through implementing FL. Audible, explicit, and traditional methods of centralized text categorization present severe privacy and security concerns because all the data is gathered at one server location. FL solves these problems by enabling the model training on multiple devices while keeping data locally. The trials have proved that the FL framework we proposed achieved a better performance than traditional centralized and domain-specific models, having an accuracy rate of 88. 4%, compared to 83. 7% and 86. 0% respectively. They suggest a 4. 7% and 2. 4% increase, respectively. In other words, the proposed framework obtained the FL model with the micro-average precision, recall, and F1-score of 88. 2%, 88. 1%, and 88. 0% respectively. The results show good generalization over virtually any domain, whether data or features. The approaches using differential privacy and secure multi-party computing achieved high accuracy rates of about 88%. 1% while Ordinary Least Squared achieved 90%, but secure multi-party computation reached the highest 88%. 2%. This proves that it is possible to balance the high accuracy of the tests and the protection of the individual's data. The evaluation of the training time of the FL framework and the computational load of the framework showed that the FL had a training time of approximately 5. There is no change in hours and inference times of 24 ms per instance that can be compared to standard models. Moreover, regarding the communication cost, the FL framework has a price of merely 10 MB per communication cycle.