Exploring Heterogeneous Federated Learning for Improved Hand Gesture Recognition

A critical aspect of human-computer interaction needs electromyography (EMG) patterns for hand gesture recognition. Hand gestures act as a medium between a person’s intention and machine control. EMG data is collected from the arm to understand hand motions and accurately predict hand gestures for increased productivity and usability in various devices. However, as EMG data is sensitive and can reveal personal information, federated learning (FL) can be used to preserve security and privacy. FL allows for the creation of a decentralized model using EMG data from multiple participants without compromising confidentiality. In this paper, we first classify data by using machine learning algorithms, k -Nearest Neighbor(KNN), Decision Tree(DT), Random Forest(RF), Gaussian Naive Bayes(GNB), Linear Discriminant Analysis(LDA), and Artificial Neural Network(ANN). The UCI2019 EMG dataset collected from the Myo Thalmic armband is utilized for training and evaluation. The data is split into $80 \%$ for training and $20 \%$ for the testing set. The best classifier with the highest accuracy for predictions is selected through various performance criteria, including accuracy, precision, F1 score, and recall. Results show that the Random Forest method performed the best, with an accuracy rate ranging from $99 \%$ to $100 \%$. In the second part of the paper, we propose preparing different models by splitting the dataset into public and private data, focusing on exploring heterogeneous federated learning. We suggest that these models can be developed using transfer learning and knowledge distillation techniques, which would help protect the privacy of the participant’s data. By incorporating each participant’s unique model into the central training model, a significant improvement in the accuracy of the hand gesture detection system can be achieved. Importantly, this approach would address intellectual privacy concerns.