Detecting sedentary behavior has increasing attention due to its significant health implications. However, distinguishing these low-intensity activities in federated learning scenarios is notably more complex than general human activity recognition. This complexity comes from heterogeneous feature distributions that can arise even for the same labeled activity, e.g., running and playing soccer, which may exhibit different sensor patterns despite both representing high-intensity activities. This paper proposes a robust personalized federated learning approach for sedentary behavior classification under adversarial conditions. Our method leverages ordinal pattern descriptors to extract meaningful symbolic representations from wearable sensor time series, then applies a meta-learning framework with Siamese Neural Networks to rapidly adapt across clients. Next, a reputation mechanism further safeguards the global model by penalizing malicious updates. Experiments on multiple public datasets show that our method achieves high F1-scores compared to baselines, affirming its ability to maintain robust performance in privacy-sensitive and adversarial environments.

This paper introduces a novel hierarchical federated learning model, Sedentary-SMELL, for classifying sedentary behavior using wearable device data. Our methodology involves transforming sensor data into Ordinal Patterns (OP) for efficient representation, training a federated autoencoder to capture standard features, and clustering users based on similar activity patterns. We employ meta-learning within clusters for enhanced pattern comparison and conclude with personalized model finetuning, adapting to individual user variations for accurate sedentary detection. Extensive testing on various datasets, including BaSA and Har UML 20, demonstrates the model’s superiority over traditional personalized methods, achieving remarkable F1scores of 0.9958 and 0.9124, respectively. Integrating a personalizing step further refines the model, tailoring it to individual user characteristics while retaining the core structure learned from meta-learning, surpassing the median performance of centralized models across all datasets.

Characterizing and monitoring patient activities through time series data is critical for identifying lifestyle patterns that may impact health outcomes. Sedentary behavior is a significant concern due to its association with various health risks. This study introduces a lightweight supervised classifier for healthcare applications based on ordinal pattern (OP) transformation to detect sedentary behavior in federated learning (FL) scenarios. Our hypothesis is grounded on the idea that sedentary behavior exhibits distinct dynamics compared to other activities, and information descriptors derived from the transformation of OPs effectively capture these differences. Next, we proceed with the FL training. We train a neural network (NN)-based encoder locally and send the local models to a server. The FL process updates the encoder weights based on the encoded representations of the clients’ data, enabling the model to learn from different participants. Finally, we personalize the model for the specific task of classifying sedentary behavior. Our approach utilizes a meta-learning framework, incorporating a Siamese NN to learn a similarity space. We fine-tune the model in this step by further training the last NN layer. This fine-tuning allows the model to adapt and specialize in accurately classifying sedentary behavior. We carry out a comprehensive analysis to support our hypothesis. We also extensively validated our proposal by comparing it with other methods over five different data sets. We obtain the best results using a smaller machine learning model compared with the best approaches in the literature. Specifically, our model has 78.73% times fewer parameters and consumes 48.67% times less energy than the best result in the literature.

Detecting transportation modes’ usability in spatiotemporal urban trajectories can provide valuable insights into the mobility preferences of urban populations, helping epidemic prevention and urban quality-of-life improvement. With this goal, we introduce POPAyI, a strategy that bases its design on the Ordinal Pattern (OP) transformation applied to mobility-related time series. POPAyI can quantify time-series dynamics with a low-complex cost, muscling time series’ characteristics without the need for high computational and methodological complexities as the current Machine Learning (ML) and Deep Learning (DL) literature. POPAyI uses polar representation and captures amplitude information in time series, bringing the multivariate capability to the standard 1D OP transformation. Our experiments show that POPAyI: (i) perfectly adapts to multi-dimensional mobility time series and natural non-linear mobility behavior. (ii) presents consistent detection results in any considered number of transportation mode’s classes with efficiency in terms of storage and computation complexity, using fewer features than ML approaches and computational resources than DL methods, e.g., reaching 10000 fewer parameters than a lightweight DL approach while increasing by 3% the F1-score.

We propose a novel deep metric learning method. Differently from many works in this area, we define a novel latent space obtained through an autoencoder. The new space, namely S-space, is divided into different regions describing positions where pairs of objects are similar/dissimilar. We locate makers to identify these regions and estimate the similarities between objects through a kernel-based Cauchy distribution to measure the markers’ distance and the new data representation. In our approach, we simultaneously estimate the markers’ position in the S-space and represent the objects in the same space. Moreover, we propose a new regularization function to prevent similar markers from collapsing altogether. Our method emphasizes the group property (separability) while preserving instance representativity. We present evidence that our proposal can represent complex spaces, for instance, when groups of similar objects are located in disjoint regions. We compare our proposal to nine different distance metric learning approaches (four of them are based on deep learning) on 28 real-world heterogeneous datasets. According to the four quantitative metrics used, our method overcomes all of the nine strategies from the literature.

In this work, we propose a solution for detecting botnet attacks on the Internet of Things (IoT) by identifying anomalies in the temporal dynamics of their devices. Given their limited computing capabilities, IoT devices are more vulnerable to attacks than conventional computers. In this scenario, botnets have a high degree of severity since they are used to trigging distributed denial-of-service attacks, which are amplified by a large number of IoT devices. Thus, solutions aiming to identify and mitigate the damage caused by botnets in IoT are urgent and essential. We evaluate the number of packets a device transmits, following a multiscale ordinal patterns transformation, and use Isolation Forest for anomaly detection. By investigating how devices evolve, we can distinguish between normal and anomalous behaviors. We apply the proposed solution to detect two major botnets for IoT: Mirai and Bashlite. We evaluated our model throughout two experimental setups. The first, using a single model for all devices, reaching 99.5% of accuracy and 99.6% of specificity, and the second, by tuning a model per device, reaching 100% of accuracy. These results show that, with the proper transformation, it is possible to use simple methods for detecting anomalies in IoT devices’ behaviors.

Federated Learning enables devices to collaboratively learn a shared prediction model while keeping all the training data on the local device and promoting clients' privacy. Vanilla federated learning models are susceptible to model poisoning attacks, where malicious nodes can inject fake models weight in order to deviate the global model to its objective. With this attack, malicious nodes intend to generate inaccurate global models or even to generate global models that make wrong inferences. This work proposes a new similarity function for federated learning applications to tackle the model poisoning vulnerability. Our method uses a new security aggregation proposal for local models based on the quantification of the heterogeneity of the data. In addition, this quantifier benefits from theoretical results found in models that use the auxiliary space proposed in our model. To assess the general behavior of our method, we evaluate our proposal in a federated scenario under non-iid data where all local models are honest, where we outperformed the vanilla model by 52.84% and 58.88% on two real-world dataset, respectively. In addition, our proposal reached 81.79% and 73.92% of F1-Score in a model poisoning experiment and outperformed the other state-of-the-art methods by 8.60% and 5.54% respectively.

This article proposes TSCLAS, a time series classification strategy for the Internet of Things (IoT) data, based on the class separability analysis of their temporal dynamics. Given the large number and incompleteness of IoT data, the use of traditional classification algorithms is not possible. Thus, we claim that solutions for IoT scenarios should avoid using raw data directly, preferring their transformation to a new domain. In the ordinal patterns domain, it is possible to capture the temporal dynamics of raw data to distinguish them. However, to be applied to this challenging scenario, TSCLAS follows a strategy for selecting the best parameters for the ordinal patterns transformation based on maximizing the class separability of the time series dynamics. We show that our method is competitive compared to other classification algorithms from the literature. Furthermore, TSCLAS is scalable concerning the length of time series and robust to the presence of missing data gaps on them. By simulating missing data gaps as long as 50% of the data, our method could beat the accuracy of the compared classification algorithms. Besides, even when losing in accuracy, TSCLAS presents lower computation times for both training and testing phases.

The traditional process of extracting knowledge from the Internet of Things (IoT) happens through Cloud Computing by offloading the data generated in the IoT device to processing in the cloud. However, this regime significantly increases data transmission and monetary costs and may have privacy issues. Therefore, it is paramount to find solutions that achieve good results and can be processed as close as possible to an IoT object. In this scenario, we developed a Neural Architecture Search (NAS) solution to generate models small enough to be deployed to IoT devices without significantly losing inference performance. We based our approach on Evolutionary Algorithms, such as Grammatical Evolution and NSGA-II. Using model size and accuracy as fitness, our proposal generated a Convolutional Neural Network model with less than 2 MB, achieving an accuracy of about 81 % in the CIFAR-10 and 99 % in MNIST, with only 150 thousand parameters approximately.

Contextual data characterize distinct regions of the city, allowing them to differentiate them according to security, entertainment, services, among others. Using contextual data to suggest routes helps to understand new aspects of a city that can change users’ perceptions of different routes. The impact of each type of contextual data may vary according to the user’s profile, which is not taken into account in most of the systems proposed by the literature. Besides, it is necessary to consider the behavior of contextual data, which changes according to the type of data. To tackle the problems mentioned above, we propose a route suggestion system with space-time risk, called GIN. The system consists of three modules, namely: identification of contextual windows, context mapping, and route personalization. Moreover, we propose a strategy to decrease the number of route requests to improve system scalability. The results show that the system adapts to sensitive changes in user’s profiles. We obtained promising by using the behavior of contextual data to avoid unnecessary requests. This strategy allowed a reduction of up to 50% of requests made to the system.

This paper investigates the precise identification of physical phenomena in the Internet of Things (IoT) context, which is one of the main challenges when dealing with the massive scale of IoT data. For this, we use information theory quantifiers in the characterization and classification of physical phenomena to minimize the effects of the lack of proper descriptions and the high heterogeneity of IoT sensors. Thus, by understanding the dynamics behind physical phenomena, we perform the classification of sensor data based on their expected behavior, not their data points. By using a simple classification algorithm, we show that the behavioral dynamics of some physical phenomena are more affected by different geographical regions than others. This gives a classification accuracy of 75% when all phenomena are considered and of 93% when considering only the invariant ones, with a worst case of false positives of 12%. This result indicates the high potential of our technique to correctly identify physical phenomena from sensor data, a fundamental issue for several applications, even in an unreliable IoT environment.

This work proposes a novel mechanism for management, orchestration and flow control in the context of the device-to-device (D2D) to deal with load balancing using the deep Q-learning (DQN) technique. To do so, we implemented a D2D network simulation environment, using the ParticiptAct dataset to evaluate the load of the cell towers in a region of Italy. The Gauss-Markov and Gilbert-Elliott models were used for mobility and packet loss, respectively, where it was considered that the towers had a disconnected coverage area, hence forming a Voronoi space. We used a Gaussian process to predict the load of the towers when they receive the packet, and a DQN to perform the balance of load of the network. This proposal presents better results than the baseline, concerning the metrics used, as well as presenting some perspectives for a future unfolding of this work.

Vehicular ad hoc network (VANET) is a fundamental building block in the design of an Intelligent Transportation System (ITS). Considering the various applications in ITS, a VANET must provide communication solutions in different situations. In particular, we are interested in dealing with situations where the unicast communication problem occurs in sparse network scenarios. In this paper, we shed light on the need for mechanisms that take into account the vehicles' trajectories. We present a characterization that shows the spatiotemporal regularity of the vehicle movement. We propose a new methodology for identifying the spatiotemporal relationship between vehicle trajectories. We create a novel method named ROSTER for unicast routing in sparse VANETs. Simulations results show that the proposed solution considerably reduces message overhead in the network by maintaining compatible levels of delivery rate in comparison with other protocols.

A social network is a valuable source of data that is useful to understand a wide range of events happening around the world, with each user being a potential contributor to accomplish this task. This work proposes a novel method to detect events in Twitter based on the calculation of entropy of the content of tweets in order to classify the most shared topic as an event or not. We observed that the entropy of the bigrams extracted from tweets are subject to a continuous phase transition when social media users start to react and interact with an event that is taking place. Hence, we propose a method to detect this phase transition, and consequently detect an event, and extract the keywords related to the corresponding event. We compared the performance of our method to other approaches of the literature and we observed that our method is the more regular among three metrics and reached the best overall performance. Furthermore, we present evidence that our method is very sensitive to correctly detect events that occur almost at same time.

In this work, we argue that Location-Based Social Media (LBSM) feeds may offer a new layer to improve traffic and transit comprehension. Initially, we showed the significant correlation between Twitter's feed and traditional traffic sensors. Then, we presented the Twitter MAPS (T-MAPS) a low-cost spatiotemporal model to improve the description of traffic conditions through tweets. T-MAPS enhance traditional traffic sensors by carrying the human lens into the transportation system. We conducted a case study by running T-MAPS and Google Maps route recommendation, in which, we showed T-MAPS viability, as an additional traffic descriptor. As a result, we noticed the median of route similarity reached 62%, and for a quarter of the evaluated trajectories, the similarity achieved between 75% and 100%. Also, we presented three route description services, based on natural language analyzes, Route Sentiment (RS), Route Information (RI), and Area' Tags (AT) aiming to enhance the route information.

It is paramount to ensure secure and trustworthy operations in Cyber-Physical Systems (CPSs), guaranteeing the integrity of sensing data, enabling access control, and safeguarding system-level operations. In this paper, we address trustworthy operations of next generation CPSs. Our idea is inspired by a trustworthy computing framework known as Proof-Carrying Code, in which foreign executables carry a model to prove that they have not been tampered with and they function as expected. In our context, we leverage the physical world--a channel that encapsulates properties impossible to tamper with remotely, such as proximity and causality--to create a challenge-response function. We call it Proof-Carrying Sensing and use it to help authenticate devices, collected data, and locations. A unique advantage of this approach, vis-à-vis traditional multi-factor or out-of-band authentication mechanisms, is that authentication proofs are embedded in sensor data and can be continuously validated over time and space without resorting to complicated cryptographic algorithms. This, in turn, makes it fit particularly well to CPSs where mobility and resource constraints are common.

Location is a fundamental service for mobile computing. Typical GPS receivers, although widely available for navigation purposes, may consume too much energy to be useful for many applications. Observing that in many sensing scenarios, the location information can be post-processed when the data is uploaded to a server, we design a cloud-offloaded GPS (CO-GPS) solution that allows a sensing device to aggressively duty-cycle its GPS receiver and log just enough raw GPS signal for post-processing. Leveraging publicly available information such as GNSS satellite ephemeris and an Earth elevation database, a cloud service can derive good quality GPS locations from a few milliseconds of raw data. Using our design of a portable sensing device platform called CLEON, we evaluate the accuracy and efficiency of the solution. Compared to more than 30 seconds of heavy signal processing on standalone GPS receivers, we can achieve three orders of magnitude lower energy consumption per location tagging.

Traffic jams frustrate drivers and cost billions per year in time and fuel consumption. In order to avoid such problems, this paper presents an intelligent transportation system that collects real-time traffic information and is able to detect and manage traffic congestion based on this information. Simulation results show that the proposed protocol can reduce the average travel time, CO2 emission and fuel consumption. In particular, the average travel time was reduced in approximately 23%, the average fuel consumption in 9% and average CO2 emission in 10%.

Vehicular ad hoc networks (VANETs) are no longer a futuristic promise but rather an attainable technology. The majority of services envisioned for VANETs either require the provisioning of multimedia support or have this support as an extremely beneficial feature. However, the highly dynamic topology of VANETs poses a demanding challenge for the fulfillment of the stringent requirements for video streaming. In this paper, we provide a deep understanding of the issue of unicast video streaming over VANETs and propose a novel protocol, VIRTUS. In video streaming, many packets are transmitted consecutively in a short period of time. VIRTUS takes this into consideration and extend the duration of the decision of nodes to forward packets from a single transmission to a time window. Furthermore, VIRTUS calculates the suitability of a node to relay packets based on a balance between geographic advancement and link stability. We also propose an extension, that separates the process of relay node selection from the transmission of video content and adopts a density-aware mechanism that adapts its behavior according to local density. Consequently, VIRTUS makes use of the reactive aspect of receiver-based solutions while remaining scalable to increases in transmission rates and density. We report through extensive realistic experiments the benefits of using VIRTUS towards delivering video at a higher quality, in a timely fashion, with lower overhead and fewer collisions.

The use of topological features, more specifically, the importance of an element related to its structural position, is a subject widely studied in the literature. For instance, the theory of complex networks provides centrality measures that have been applied to a large variety of fields (e.g., social sciences and biology). In this work, we propose a new topological measure, the Sink Betweenness (SBet), which stems from the theory of complex networks but is adapted to Wireless Sensor Networks (WSNs) to capture relevant information for this kind of network. We also provide a distributed algorithm to calculate it, and show its applicability to two different scenarios. The first one is focused on data fusion applications for event-driven WSNs, where we devise a tree-based data collection algorithm that takes advantage of node centrality to improve the data fusion efficiency. The second scenario is focused on energy balancing problems, more specifically in a problem called energy hole , where nodes closer to the sink are more likely to relay a larger number of packets than those that are further. This phenomenon is strongly related to the topology induced by the deployment of nodes along the sensor field, and it can be effectively captured by the SBet metric. Thus, we devise a data collection algorithm that is able to distribute the relay task more evenly. Simulation results show that the SBet metric can be satisfactorily used in both scenarios. We compare the proposed approach with some of the most efficient available data fusion algorithms, and show that the proposed algorithm generates consistently good-quality data collection infrastructures which require significantly smaller overhead. The use of SBet allows to alleviate the energy-hole effects by evenly balancing the relay load, and thus increasing the network lifetime. These two applications illustrate how the topology awareness can be used to improve different network functions in a WSN.

Large scale dense Wireless Sensor Networks (WSNs) will be increasingly deployed in different classes of applications for accurate monitoring. Due to the high density of nodes in these networks, it is likely that redundant data will be detected by nearby nodes when sensing an event. Since energy conservation is a key issue in WSNs, data fusion and aggregation should be exploited in order to save energy. In this case, redundant data can be aggregated at intermediate nodes reducing the size and number of exchanged messages and, thus, decreasing communication costs and energy consumption. In this work, we propose a novel Data Routing for In-Network Aggregation, called DRINA, that has some key aspects such as a reduced number of messages for setting up a routing tree, maximized number of overlapping routes, high aggregation rate, and reliable data aggregation and transmission. The proposed DRINA algorithm was extensively compared to two other known solutions: the Information Fusion-based Role Assignment (InFRA) and Shortest Path Tree (SPT) algorithms. Our results indicate clearly that the routing tree built by DRINA provides the best aggregation quality when compared to these other algorithms. The obtained results show that our proposed solution outperforms these solutions in different scenarios and in different key aspects required by WSNs.

Video streaming capabilities over Vehicular Ad Hoc Networks (VANETs) are crucial to the development of interesting and valuable services. However, VANETs are a challenging environment to this kind of communication due to the dispersion and movement of vehicles. In this work, we present a feasible solution to this problem. The VIdeo Reactive Tracking-based UnicaSt protocol (VIRTUS) is a receiving-based solution that uses vehicles' current and future location for a selection policy of relaying nodes. It fulfills video streaming requirements without incurring into an excessive number of transmissions. Besides that, it outperforms other baseline solutions.

Transmitting video content over Vehicular Ad Hoc Networks (VANETs) faces a great number of challenges caused by strict QoS (Quality of Service) requirements and highly dynamic network topology. In order to tackle these challenges, multipath forwarding schemes can be regarded as potential solutions. However, route coupling will severely impair the performance of multipath schemes. In this work, we present a LocatIon-Aware multIpaTH videO streamiNg (LIAITHON) scheme to address video streaming over urban VANETs. LIAITHON uses location information to discover two relatively short paths with minimum route coupling effect. The performance results have shown it outperforms the underlying single path solution as well as the node-disjoint multipath solution.

Location is a fundamental service for mobile computing. Typical GPS receivers, although widely available, consume too much energy to be useful for many applications. Observing that in many sensing scenarios, the location information can be post-processed when the data is uploaded to a server, we design a Cloud-Offloaded GPS (CO-GPS) solution that allows a sensing device to aggressively duty-cycle its GPS receiver and log just enough raw GPS signal for post-processing. Leveraging publicly available information such as GNSS satellite ephemeris and an Earth elevation database, a cloud service can derive good quality GPS locations from a few milliseconds of raw data. Using our design of a portable sensing device platform called CLEO, we evaluate the accuracy and efficiency of the solution. Compared to more than 30 seconds of heavy signal processing on standalone GPS receivers, we can achieve three orders of magnitude lower energy consumption per location tagging.

Trajectory-based services require continuous user location sensing. GPS is the most common outdoor location sensor on mobile devices. However, the high energy consumption of GPS sensing prohibits it to be used continuously in many applications. In this paper, we propose a Low Energy Assisted Positioning (LEAP) solution that carefully partitions the GPS signal processing pipeline and shifts delay tolerant position calculations to the cloud. The GPS receiver only needs to be on for less than a second to collect the sub-millisecond level propagation delay for each satellites signal. With a reference to a nearby object, such as a cell tower, the LEAP server can infer the rest of the information necessary to perform GPS position calculation. We analyze the accuracy and energy benefit of LEAP and use real user traces to show that LEAP can save up to 80% GPS energy consumption in typical trajectory-based service scenarios.

This paper presents a quantitative comparison of the performance of skin detection techniques for a Web-based system. The procedures under analysis employ dimensionality reduction, with or without color correction. We propose a rule for deciding if an image requires color correction and, if needed, which is the best procedure. Geometric rules, parametric methods and histogram-based techniques are compared in terms of accuracy and performance. With the proposed heuristic for color correction, the best performances are achieved by a classifier based on the mixture of two Gaussian laws applied to data projected with principal component analysis and a by a KDD-based rule that operates on the RGB space.