A privacy-preserving distributed transfer learning in activity recognition

Along with the widespread use of smartphones, activity recognition using embedded inertial sensors has intrigued researchers. The learning and employing activity recognition systems suffers from some challenges. Training data of different persons should be involved in the learning procedure. In contrast, the activities of each individual should remain private. In addition, the style of different individuals activities makes the training phase become a challenging problem. This paper presents two privacy-preserving transfer learning algorithms for activity recognition application namely as \(P^{2}S^{2}{} { TM}\) and \(P^{2}{} { DS}^{2}{} { TM}\). The \(P^{2}S^{2}{} { TM}\) algorithm is designed based on preserving data privacy by the client, while the \(P^{2}{} { DS}^{2}{} { TM}\) algorithm trusts the preservation of the data privacy by both the client and the server. The performance of the proposed algorithms is influenced by a random matrix. The performances of the algorithms are compared with those of the classic PPSVM classifier for different dimensions of the random matrix. The evaluation of the proposed algorithms revealed superior performance comparing with PPSVM in different scenarios. For the random matrix with the dimension of \(500\times 561\), The \(P^{2}S^{2}{} { TM}\) and \(P^{2}{} { DS}^{2}{} { TM}\) algorithms are able to relatively reduce privacy-preserving activity recognition error rate by 68% and 16% respectively; while preserving privacy with an acceptable rate.