While applying deep learning models has significantly enhanced the performance of sensor-based human activity recognition (HAR), their deployment on resource-limited mobile and wearable devices presents challenges. While reducing model size is an alternative, it often leads to performance degradation. This study introduces the integration of attention mechanisms and knowledge distillation techniques to enhance the recognition performance of lightweight sensor-based human activity recognition models. Employing three activity recognition datasets (Opportunity, Wisdm, and Sensors), initial experiments investigate the individual effects of response-based knowledge distillation and attention mechanisms on the performance of the lightweight model. Results demonstrate that attention surpasses distillation in recognition success. To further improve performance, we introduce two combined approaches: response and attention-based distillation and response-based distillation with attention directly applied to the lightweight model. Both approaches outperform attention alone, while the latter, by directly incorporating attention into the student without requiring an attention-based teacher, eliminates the need for a pre-trained attention-based teacher. Evaluating resource consumption, the performance-boosted lightweight model is compared to a moderate-size model, requiring approximately 3-4 times more flops and parameters. The improved lightweight model outperformed the moderate-size model across all datasets regarding recognition performance.

The traditional method to authenticate users on mobile devices or applications requires usernames and passwords. This approach authenticates the user only at the entry point of an application without providing continuous authentication during the whole session. This paper explores the use of behavioral biometrics, which involves tracking the unique movements of a user while interacting with a device, for continuous authentication on a mobile banking application. As a methodology, we use binary classification and explore the performance of deep learning algorithms. A dataset is collected from 45 participants using a mobile banking application in Turkey. We train four different types of deep architectures, including Multilayer Perceptron (MLP), LSTM, bi-directional LSTM, and convolutional LSTM. The dataset includes data from both touch screens and motion sensors. The results of the experiments reveal that MLP and the convolutional-LSTM algorithms achieve the best performance on raw data from both motion sensors and touch screens. Accuracy rates are over 99.85%, and FAR, FRR, and EER are below 0.5%.

In this paper, we use Hand Movement Orientation and Grasp (HMOG) sensor data to authenticate smart phone users. The way a user holds, grasps a mobile phone or touches to it are all key factors for authentication. At the moment of a user makes an event on his/her smart phone, three sensors automatically collect data about magnitude, angular speed and acceleration. Moreover, touching and holding events also produce data about pressure and coordinates. In this paper, we build four types of machine learning algorithms (decision forest, boosted decision tree, support vector machine, logistic regression) to predict user authentication. The data used in this experiment (HMOG) are collected from 100 students. We compare the results of the algorithms and for our scenario, we show that decision forest algorithm with normalized data gives best results.

Most of the studies on human activity recognition using smartphones and smartwatches are performed in an offline manner. In such studies, collected data is analyzed in machine learning tools with less focus on the resource consumption of these devices for running an activity recognition system. In this paper, we analyze the resource consumption of human activity recognition on both smartphones and smartwatches, considering six different classifiers, three different sensors, different sampling rates and window sizes. We study the CPU, memory and battery usage with different parameters, where the smartphone is used to recognize seven physical activities and the smartwatch is used to recognize smoking activity. As a result of this analysis, we report that classification function takes a very small amount of CPU time out of total app's CPU time while sensing and feature calculation consume most of it. When an additional sensor is used besides an accelerometer, such as gyroscope, CPU usage increases significantly. Analysis results also show that increasing the window size reduces the resource consumption more than reducing the sampling rate. As a final remark, we observe that a more complex model using only the accelerometer is a better option than using a simple model with both accelerometer and gyroscope when resource usage is to be reduced.

Smoking is known to be one of the main causes for premature deaths. A reliable smoking detection method can enable applications for an insight into a user's smoking behaviour and for use in smoking cessation programs. However, it is difficult to accurately detect smoking because it can be performed in various postures or in combination with other activities, it is less-repetitive, and it may be confused with other similar activities, such as drinking and eating. In this paper, we propose to use a two-layer hierarchical smoking detection algorithm (HLSDA) that uses a classifier at the first layer, followed by a lazy context-rule-based correction method that utilizes neighbouring segments to improve the detection. We evaluated our algorithm on a dataset of 45 hours collected over a three month period where 11 participants performed 17 hours (230 cigarettes) of smoking while sitting, standing, walking, and in a group conversation. The rest of 28 hours consists of other similar activities, such as eating, and drinking. We show that our algorithm improves recall as well as precision for smoking compared to a single layer classification approach. For smoking activity, we achieve an F-measure of 90-97% in person-dependent evaluations and 83-94% in person-independent evaluations. In most cases, our algorithm corrects up to 50% of the misclassified smoking segments. Our algorithm also improves the detection of eating and drinking in a similar way. We make our dataset and data logger publicly available for the reproducibility of our work.

The exponential growth in mobile data exchange rates, which is further augmented by the rapid proliferation of mobile devices, significantly increases the capital and operational expenses of the cellular network operators. Growing expenditures combined with environmental awareness have driven the network operators to provide more energy-efficient services and reduce their carbon dioxide footprint by using green networking methods. In this paper, we focus on green dynamic base station (BS) topology management according to the varying traffic conditions for the next-generation heterogenous cellular networks. We first create a detailed 3-D map of a selected pilot application area in order to obtain a realistic spatio-temporal user density estimation. Then, we formulate a novel optimization problem to find the best possible topology, which minimizes the overall network power consumption while satisfying quality-of-service requirements of the subscribers. As an additional problem, we derive a pico BS deployment methodology on top of the existing network infrastructure to accommodate the peak traffic conditions. Finally, we propose offline-centralized and online-distributed green dynamic BS operation algorithms for power saving during the operation phase. Extensive simulations based on a realistic and a random test case scenario show that our dynamic BS topology management methods achieve significant power savings compared with static operation.

We investigate the following fundamental question—how fast can information be collected from a wireless sensor network organized as tree? To address this, we explore and evaluate a number of different techniques using realistic simulation models under the many-to-one communication paradigm known as convergecast. We first consider time scheduling on a single frequency channel with the aim of minimizing the number of time slots required (schedule length) to complete a convergecast. Next, we combine scheduling with transmission power control to mitigate the effects of interference, and show that while power control helps in reducing the schedule length under a single frequency, scheduling transmissions using multiple frequencies is more efficient. We give lower bounds on the schedule length when interference is completely eliminated, and propose algorithms that achieve these bounds. We also evaluate the performance of various channel assignment methods and find empirically that for moderate size networks of about 100 nodes, the use of multifrequency scheduling can suffice to eliminate most of the interference. Then, the data collection rate no longer remains limited by interference but by the topology of the routing tree. To this end, we construct degree-constrained spanning trees and capacitated minimal spanning trees, and show significant improvement in scheduling performance over different deployment densities. Lastly, we evaluate the impact of different interference and channel models on the schedule length.

We investigate the tradeoff between two mutually conflicting performance objectives-throughput and delay-for fast, periodic data collection in tree-based sensor networks arbitrarily deployed in 2-D. Two primary factors that affect the data collection rate (throughput) and timeliness (delay) are: 1) efficiency of the link scheduling protocol, and 2) structure of the routing tree in terms of its node degrees and radius. In this paper, we utilize multiple frequency channels and design an efficient link scheduling protocol that gives a constant factor approximation on the optimal throughput in delivering aggregated data from all the nodes to the sink. To minimize the maximum delay subject to a given throughput bound, we also design an (α, β)-bicriteria approximation algorithm to construct a Bounded-Degree Minimum-Radius Spanning Tree, with the radius of the tree at most β times the minimum possible radius for a given degree bound Δ*, and the degree of any node at most Δ*+ α , where α and β are positive constants. Lastly, we evaluate the efficiency of our algorithms on different types of spanning trees and show that multichannel scheduling, combined with optimal routing topologies, can achieve the best of both worlds in terms of maximizing the aggregated data collection rate and minimizing the maximum packet delay.

Falls are identified as a major health risk for the elderly and a major obstacle to independent living. Considering the remarkable increase in the elderly population of developed countries, methods for fall detection have been a recent active area of research. However, existing methods often use only wearable sensors, such as acceloremeters, or cameras to detect falls. In this demonstration, in contrast to the state of the art solutions, we focus on the use of multi-modal wireless sensor networks within the WeCare framework. WeCare system is developed as a solution for independent living applications by remotely monitoring the health and well-being of its users. We describe the general structure of WeCare and demonstrate its fall detection method. Our set-up not only includes scalar sensors to detect falls and motion but also consists of embedded cameras and RFID tags and uses sensor fusion techniques to improve the success of fall detection and minimize the false positives.

Fast and periodic collection of aggregated data is of considerable interest for mission-critical and continuous monitoring applications in sensor networks. In the many-to-one communication paradigm known as convergecast, we consider scenarios where data packets are aggregated at each hop en route to a sink node along a tree-based routing topology and focus on maximizing the data collection rate at the sink by employing TDMA scheduling and multiple frequency channels. Our key result in the paper lies in proving that minimizing the schedule length for an arbitrary network in the presence of multiple frequencies is NP-hard, and in designing approximation algorithms with worst-case provable performance guarantees for geometric networks. In particular, we design a constant factor approximation for networks modeled as unit disk graphs (UDG) where every node has a uniform transmission range, and a O(Delta(T)log n) approximation for general disk graphs where nodes have different transmission ranges; n is the number of nodes in the network and Delta(T) is the maximum node degree on a given routing tree T. We also prove that a constant factor approximation is achievable on UDG even for unknown routing topologies so long as the maximum node degree in the tree is bounded by a constant. We also show that finding the minimum number of frequencies required to remove all the interfering links in an arbitrary network in NP-hard. We give an upper bound on the maximum number of such frequencies required and propose a polynomial time algorithm that minimizes the schedule length under this scenario. Finally, we evaluate our algorithms through simulations and show various trends in performance for different network parameters.

What is the fastest rate at which we can collect a stream of aggregated data from a set of wireless sensors organized as a tree? We explore a hierarchy of techniques using realistic simulation models to address this question. We begin by considering TDMA scheduling on a single channel, reducing the original problem to minimizing the number of time slots needed to schedule each link of the aggregation tree. The second technique is to combine the scheduling with transmission power control to reduce the effects of interference. To better cope with interference, we then study the impact of utilizing multiple frequency channels by introducing a simple receiver-based frequency and time scheduling approach. We find that for networks of about a hundred nodes, the use of multi-frequency scheduling can suffice to eliminate most of the interference. The data collection rate then becomes limited not by interference, but by the maximum degree of the routing tree. Therefore we consider finally how the data collection rate can be further enhanced by the use of degree-constrained routing trees. Considering deployments at different densities, we show that these enhancements can improve the streaming aggregated data collection by as much as 10 times compared to the baseline of single-channel data collection over non-degree constrained routing trees. Addition to our primary conclusion, in the frequency scheduling domain we evaluate the impact of different interference models on the scheduling performance and give topology-specific bounds on time slot and frequency channel requirements.

This paper presents measurements of radio interference using "ambient munode" sensor nodes. By varying distances and frequencies we get a measure of the interference caused by transmissions on adjacent bands. Our observations show that adjacent spectrum interference influences the data delivery, considerably. Channels should be separated in the spatial or in the frequency domains if interference is to be avoided. In addition, the distance to simultaneous transmitters and the number of simultaneous transmissions are highly correlated with channel spacing. Therefore, channel spacing can be adjusted according to spatial distances so that multiple concurrent transmissions can be performed without interference. We also give proposals for further investigation on the usage of this correlation that are relevant to the design of future multi-channel protocols