Decentralized CNN Training for Table Tennis Serve Classification Using Federated Learning

This paper provides an original FL CNN framework specially designed for categorizing serve shots in table tennis. The approach deals with data security protection and stability challenges in the model. Standard pattern recognition methods typically require the collection of data to be centralized, which is quite problematic in terms of privacy and limits the number of available datasets. The FL method is applied as a collaborative approach to training models with distributed datasets where privacy is maintained while the set datasets are augmented to improve the resulting model. Comparison was also made of the FL CNN architecture with that of a conventional centralized structure. The FL model reached a classification accuracy of 92.5%, Precision of 93.2%, and Recall of 91. 8%, and the F1-score was equal to 92. 5%. On the other hand, the accuracy rate of the centralized model I achieved was 89. 70% for centrepiece agreement and a precision rate of 90. 5%, a recall rate of 88.Therefore, the percentage increases by 5%, and the median and metamorphotic increases by 19%. Specifically, it achieves an accuracy of 92%, Precision of 94%, Recall of 93% and F1-score of 89.3%. The study shows that the FL technique effectively controls many serve shots, and data privacy is well protected. The investigation of training loss revealed that the FL model had a lessened training loss compared to the centralized model over 20 epochs of training. The training losses in the FL model and the centralized model were 0 at the end, 22 and 18, respectively.