Smart devices, such as smartphones, wearables, robots, and others, can collect vast amounts of data from their environment. This data is suitable for training machine learning models, which can significantly improve their behavior, and therefore, the user experience. Federated learning is a young and popular framework that allows multiple distributed devices to train deep learning models collaboratively while preserving data privacy. Nevertheless, this approach may not be optimal for scenarios where data distribution is non-identical among the participants or changes over time, causing what is known asconcept drift. Little research has yet been done in this field, but this kind of situation is quite frequent in real life and poses new challenges to both continual and federated learning. Therefore, in this work, we present a new method, called Concept-Drift-Aware Federated Averaging (CDA-FedAvg). Our proposal is an extension of the most popular federated algorithm, Federated Averaging (FedAvg), enhancing it for continual adaptation under concept drift. We empirically demonstrate the weaknesses of regular FedAvg and prove that CDA-FedAvg outperforms it in this type of scenario.

Federated learning is the state-of-the-art paradigm for training a learning model collaboratively across multiple distributed devices while ensuring data privacy. Under this framework, different algorithms have been developed in recent years and have been successfully applied to real use cases. The vast majority of work in federated learning assumes static datasets and relies on the use of deep neural networks. However, in real-world problems, it is common to have a continual data stream, which may be non-stationary, leading to phenomena such as concept drift. Besides, there are many multi-device applications where other, non-deep strategies are more suitable, due to their simplicity, explainability, or generalizability, among other reasons. In this paper we present Ensemble and Continual Federated Learning, a federated architecture based on ensemble techniques for solving continual classification tasks. We propose the global federated model to be an ensemble, consisting of several independent learners, which are locally trained. Thus, we enable a flexible aggregation of heterogeneous client models, which may differ in size, structure, or even algorithmic family. This ensemble-based approach, together with drift detection and adaptation mechanisms, also allows for continual adaptation in situations where data distribution changes over time. In order to test our proposal and illustrate how it works, we have evaluated it in different tasks related to human activity recognition using smartphones.

Presently, smartphones are used more and more for purposes that have nothing to do with phone calls or simple data transfers. One example is the recognition of human activity, which is relevant information for many applications in the domains of medical diagnosis, elderly assistance, indoor localization, and navigation. The information captured by the inertial sensors of the phone (accelerometer, gyroscope, and magnetometer) can be analyzed to determine the activity performed by the person who is carrying the device, in particular in the activity of walking. Nevertheless, the development of a standalone application able to detect the walking activity starting only from the data provided by these inertial sensors is a complex task. This complexity lies in the hardware disparity, noise on data, and mostly the many movements that the smartphone can experience and which have nothing to do with the physical displacement of the owner. In this work, we explore and compare several approaches for identifying the walking activity. We categorize them into two main groups: the first one uses features extracted from the inertial data, whereas the second one analyzes the characteristic shape of the time series made up of the sensors readings. Due to the lack of public datasets of inertial data from smartphones for the recognition of human activity under no constraints, we collected data from 77 different people who were not connected to this research. Using this dataset, which we published online, we performed an extensive experimental validation and comparison of our proposals.

Mobile phones are increasingly used for purposes that have nothing to do with phone calls or simple data transfers, and one such use is indoor inertial navigation. Nevertheless, the development of a standalone application able to detect the displacement of the user starting only from the data provided by the most common inertial sensors in the mobile phones (accelerometer, gyroscope and magnetometer), is a complex task. This complexity lies in the hardware disparity, noise on data, and mostly the many movements that the mobile phone can experience and which have nothing to do with the physical displacement of the owner. In our case, we describe a proposal, which, after using quaternions and a Kalman filter to project the sensors readings into an Earth Centered inertial reference system, combines a classic Peak-valley detector with an ensemble of SVMs (Support Vector Machines) and a standard deviation based classifier. Our proposal is able to identify and filter out those segments of signal that do not correspond to the behavior of “walking”, and thus achieve a robust detection of the physical displacement and counting of steps. We have performed an extensive experimental validation of our proposal using a dataset with 140 records obtained from 75 different people who were not connected to this research.

Most facial verification methods assume that training and testing sets contain independent and identically distributed samples, although, in many real applications, this assumption does not hold. Whenever gathering a representative dataset in the target domain is unfeasible, it is necessary to choose one of the already available (source domain) datasets. Here, a study was performed over the differences among six public datasets, and how this impacts on the performance of the learned methods. In the considered scenario of mobile devices, the individual of interest is enrolled using a few facial images taken in the operational domain, while training impostors are drawn from one of the public available datasets. This work tried to shed light on the inherent differences among the datasets, and potential harms that should be considered when they are combined for training and testing. Results indicate that a drop in performance occurs whenever training and testing are done on different datasets compared to the case of using the same dataset in both phases. However, the decay strongly depends on the kind of features. Besides, the representation of samples in the feature space reveals insights into what extent bias is an endogenous or an exogenous factor.