Federated Deep Learning for a Privacy-preserved Activity Recognition

Background Electrical line workers are among the group of workers who have one of the top 10 most dangerous jobs in America [1]. Monitoring workers' activities and specifying their activities is one of the fundamental steps in taking any preventive measures towards providing a safe working environment for these workers [2]. To monitor the safety of workers and prevent injuries, it is essential to propose a framework which does not interfere with their daily tasks. Recent advances in wearable sensors are promising in obtaining this objective, however, the concern about privacy of collected data remains an important issue, preventing their mass adoption in industry [3,4]. The information collected from sensors could vary from sensitive physiological data like heart rate to other types of data such as motion or sound. On the other hand, regardless of the type of data used, the notion of wearing sensors, enabling any type of monitoring, causes hesitance to adopt them by workers. Therefore, addressing this issue and ensuring data confidentiality could facilitate better ergonomic interventions at workplaces with less resistance from workers. It’s not long since data confidentiality has been a pressing issue in the era of artificial intelligence, and various approaches have been proposed to prevent data breaches. Yet, there is still a gap in guaranteeing the data privacy in ergonomics and safety contexts. In this paper, we developed a decentralized learning method which addresses this gap. Approach In this study, participants performed 10 main activities of electrical line workers in a lab environment while their acceleration data was collected using a wearable sensor on their dominant wrist. Time, frequency, and wavelet (time and frequency) features were extracted from the filtered accelerometer’s signals as inputs to the model. In this study, we use the federated weighted averaging with adaptive learning rate method to classify the activities of individuals using these features. Our model consisted of a LSTM layer, trained locally on participants' wearables. After each step of local training, the updated weights were iteratively broadcasted to a server to update the global weights. This step ensures data confidentiality, since unlike the centralized models, the training data are not combined on the server, instead only the weights are aggregated and updated. After training, we evaluated the performance of our model to that of frequently used centralized machine learning models such as LSTM, k-nearest neighbors, support vector machine, and random forest [5]. Results Our results showed