Human Activity Recognition based on Noval Ensemble CNN Classifier with Privacy-Preserved Knowledge Graph-based Feature Extraction

Abstract

Human Activity Recognition (HAR) is crucial in various applications. The dual challenge of preserving data privacy while attaining accurate classification poses a major challenge in the field of HAR where there is an abundance of sensitive information. This paper introduces a novel methodology called Privacy-Preserving Graph Convolutional Network Ensemble (PriGCN-ECNN) Classifier. The approach leverages a concept known as Deep Autoencoder-like Non-Negative Matrix Factorization (DANMF) for privacy preservation. To elevate the feature extraction in HAR, both spatial and temporal features are adopted in this study with the concept of GCN. In HAR, where activities highly correlate, GCN works better than standard models. Finally, the work introduces ECNN, an ensemble model that uses weighted averages for CNN model variants. The suggested model is tested on UniMiB-SHAR, MotionSense, and WISDM Actitracker utilizing accuracy, F1-score, and confusion matrix. The model improves accuracy to 96.01%, 99.2% in Motion sense data, and 97.4% in WISDM Actitracker.  The proposed model exhibits an enhancement in accuracy (0.4%) as compared to the CNN-based methodology employed in the UniMiB-SHAR dataset. Hence, the Pri-GCN-ECNN not only guarantees the preservation of data privacy but also enhances the accuracy and reliability of HAR.