Balancing Activity Recognition and Privacy Preservation with a Multi-objective Evolutionary Algorithm

With the widespread of miniaturized inertial sensors embedded in wearable devices, an increasing number of individuals monitor their daily life activities through consumer electronic products. However, long-lasting data collection (e.g., from accelerometer) may expose the users to privacy violations, such as the leakage of personal details. To help mitigate these aspects, we propose an approach to conceal subject’s personal attributes (i.e., gender) while maximizing the accuracy on both the monitoring and recognition of human activity. In particular, a Multi-Objective Evolutionary Algorithm (MOEA), namely the Non-dominated Sorting Genetic Algorithm II (NSGA-II), is applied to properly weight input features extracted from the raw accelerometer data acquired with a wrist-worn device (Empatica E4). Experiments were conducted on a large-scale and real life dataset, and validated by adopting the Random Forest algorithm with 10-fold cross validation. Findings demonstrate that the proposed method can highly limit gender recognition (from 89.37% using all the features to 64.38% after applying the MOEA algorithm) while only reducing the accuracy of activity recognition by 5.45% points (from 89.59% to 84.14%).