Federated Learning (FL) is susceptible to adversarial attacks, such as Label Flipping (LF) and Backdoor, where malicious clients manipulate the updates of their local model to reduce the global model’s performance. Traditional FL relies on a centralised aggregator, which must be trusted, creating a single point of failure. This centralization not only increases computational cost but also introduces scalability challenges. To address these issues, we propose a Blockchain (BC) based FL framework that decentralises the aggregation process and incorporates an enhanced Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method to identify and remove malicious updates without ignoring minor groups. Our approach eliminates the need for a centralised aggregator by leveraging BC’s smart contracts to aggregate the global model. Simultaneously, Enhanced DBSCAN identifies malicious updates in the parameter space, effectively mitigating adversarial influence while preserving privacy. We evaluate both of our framework and the traditional FL under LF red and Backdoor attacks, experimental results demonstrate that our approach outperforms the traditional FL according to multiple metrics, including accuracy, loss, precision, recall, and F1-score. These findings emphasise the effectiveness of our BC-based decentralised aggregation combined with enhanced DBSCAN technique in improving the robustness and security of FL systems.

We present FedADHD, a distributed Federated Learning (FL) framework designed to enhance the diagnosis of Attention Deficit Hyperactivity Disorder (ADHD) while preserving patient privacy. Leveraging the HYPERAKTIV dataset, which contains health activity and neuropsychological data from adults diagnosed with ADHD, we developed a hybrid deep learning model that integrates Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. The model analyzes key diagnostic indicators such as the ADHD Confidence Index, omission/commission T-scores, and reaction times extracted from the Conners CPT-II. FedADHD enables collaborative training across multiple mental healthcare institutions, treated as autonomous agents in a federated network, without requiring raw data sharing. This privacy-preserving architecture addresses critical ethical and legal concerns in mental health research. Evaluated using Accuracy, Precision, Recall, F1-score, and Matthews Correlation Coefficient (MCC), FedADHD outperforms traditional centralized Machine Learning (ML) models, showcasing its robustness and generalization capabilities. This work demonstrates how distributed ML and secure collaboration can significantly advance mental health diagnostics in real-world, multi-institutional settings.

Mobile behavioural biometrics, leveraging touchscreen and background sensor data, offer a promising approach to Continuous Authentication (CA). However, the performance of these systems can vary significantly under different attack scenarios. This study evaluates the effectiveness of the regularisation technique in improving authentication accuracy within Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) architecture. Using the BehavePassDB dataset, we test four regularisation techniques (Ridge, Lasso, Bayesian, and ElsticNet) on accelerometer sensor data across various tasks, including Keystroke, Readtext, Gallery, and Tap. Results demonstrate that integrating regularisation techniques with LSTM-based models consistently outperforms the BBCA system, particularly in random and skilled attack scenarios, with Area Under the Curve (AUC) improvements of up to 15%. These findings underscore the potential of combining advanced neural networks with regularisation techniques to enhance mobile biometric systems.

The Controller Area Network (CAN) protocol is the main communication backbone in modern vehicles, enabling data exchange between Electronic Control Units (ECUs). However, its lack of built-in security mechanisms exposes Connected Vehicles (CV) to a growing range of cyber threats. To address this, we propose a distributed intrusion detection framework that integrates Model-Agnostic Meta-Learning (MAML) with Federated Learning (FL), offering adaptability and privacy preservation. The system leverages an LSTM-based model to learn temporal patterns in CAN message IDs, enabling the detection of known and previously unseen attacks. MAML enhances the model's generalization by facilitating rapid adaptation to new threat types using limited data, while FL enables collaborative training across multiple CVs without sharing raw data. Our distributed approach ensures continuous learning and robustness in real-world automotive environments. Experimental results demonstrate the framework's effectiveness in detecting complex intrusions while maintaining data privacy across participating vehicles.

Mobile behavioural biometrics, leveraging touchscreen and background sensor data, have emerged as a promising solution for Continuous Authentication (CA) on mobile devices, enabling secure user authentication. In this paper, we introduce an enhanced CA framework named AI-MBBCA, that integrates a Genetic Algorithm (GA) for optimal training hyper-parameter selection and an Isolation Forest (IF) as a secondary layer for impostor attack detection. A hybrid Long Short-Term Memory (LSTM) network, trained using a triplet loss function and augmented with a regularisation method, effectively captures spatial and temporal patterns in user behaviour. Experimental evaluations on the BehavePassDB dataset demonstrate that AI-MBBCA significantly improves authentication accuracy and reduces error rates across multiple tasks, with notable improvements in the Area Under the Curve (AUC) compared to two other approaches from the literature. Integrating AI-Driven optimisation, including GA and IF-based anomaly detection, paves the way for more resilient and adaptive Behavioural Biometrics Continuous Authentication (BBCA) systems, addressing the challenges posed by sophisticated forgery scenarios in dynamic mobile environments.

Federated Learning (FL) is a powerful Machine Learning (ML) technique that allows multiple clients to collaborate on training models while keeping their data private. Unlike traditional centralized methods, FL ensures that data are kept separate, which helps to protect privacy. However, an important area that needs more research while using the FL system is detecting harmful models within the Internet of Things (IoT) context. For example, poisoning attacks, where compromised clients introduce harmful data, can degrade the model’s overall performance or lead to incorrect predictions. This paper comprehensively reviews of recent attacks in FL within IoT networks, along with defense mechanisms and common FL frameworks. It begins by highlighting the significance of FL in IoT networks, exploring its applications, benefits, and inherent security challenges. It then explores specific attacks targeting FL in IoT networks. The defensive strategies are evaluated, including their performance metrics, datasets used, and related work, providing a comparative analysis of these techniques. Common FL frameworks and their criteria are reviewed. Our goal is to offer a detailed understanding and solutions to enhance the strength and resilience of FL systems in IoT networks.

Recently, Session Initiation Protocol (SIP) involved as a core communication protocol in several technologies such as IMS, VoLTE and integrated with other applications. SIP known to be highly efficient protocol for providing high multimedia communication (voice and video) in industry and research society. In production, SIP service providers support millions of users and manage thousands of concurrent call-requests efficiently via a wide scale of cluster SIP servers. However, asynchronous processes-load among cluster nodes and weak resource utilization considered to be a serious issue in SIP technology, especially with high call-requests distribution. This paper proposes a dispatcher mechanism called SIP- Cluster Load Synchronization (SIP-CLS) which distributes SIP call-requests in cluster based on processes load utilization for efficient resource allocation. The mechanism takes advantage of NoSQL cache system to share and update cluster nodes load status. Evaluation have been conducted by torturing SIP server with high traffic in two different scenarios. As a results, the proposed mechanism performs efficiently compared to other load balancing mechanism by providing high resource utilization in term of bandwidth and CPU.

Federated Learning (FL) is a technique used in Internet of Things (IoT) networks to enhance data privacy through decentralised Machine Learning (ML). However FL faces challenges due to the Non-Independent and Identically Distributed (Non-IID) data that is stored on various devices. Each device typically has a unique Non-IID subset of data from its local environment. This Non-IID distribution can be manipulated by poisoning attacks, where malicious modifications disrupt the global model. To addresses these complex in both IID and Non-IID data environments, we introduce AM2DN-FL. This adaptive approach identifies and removes malicious models in FL system, using a dual-sided defense strategy that leverages server and client components to combat Label-Flipping (LF) and backdoor attacks. AM2DN-FL employs an refined Local Outlier Factor (LOF) algorithm with an adaptive threshold based on Genetic Algorithms (GA) to fine-tuning the optimal threshold selection. Our simulation outcomes, utilizing the MNIST and CIFAR10 datasets for IID and Non-IID scenarios, demonstrate that our innovative approach outperforms other previously examined approaches in the literature across various performance metrics, such as Accuracy Rate (ACC), Attack Success Rate (ASR), Recall, Precision, and CPU run-time.

Federated Learning (FL) is a crucial technology in decentralized Machine Learning (ML), prominently used within Internet of Things (IoT) networks to enhance data privacy. However, it is threatened by poisoning attacks, where harmful data alterations can significantly disrupt learning processes. This paper introduces a novel solution, Entropy-based Robust Defense Federated Learning (ERDFL), to counteract these disruptions. Our approach leverages entropy information for enhanced detection of malicious models and also innovatively adjusts detection thresholds in real-time, thereby effectively identifying and excluding potentially malicious clients within the FL process. Our simulation results, using the Mnist, Fashion-Mnist, and IMDB datasets, demonstrate that our novel approach surpasses other previously studied approaches in the literature across multiple performance metrics, including Accuracy Rate (ACC), Attack Success Rate(ASR), Loss Rate (LR) and CPU aggregation run-time.

Federated Learning (FL) is rapidly growing in popularity as a decentralized approach and is being adopted in smart building systems and energy forecasting without accessing sensitive data. Specifically, clients train their models using their own data. After that, only their model parameters are sent to the central server, which aggregates them by averaging the weights and then sends back the newly formed model to each client. However, challenges arise when dealing with heterogeneous multivariate time-series data with different distributions. This leads to higher-performing clients contributing to the global update more than the others, and slower convergence where the global model takes more time to generalize across the clients. In this paper, we propose an enhanced aggregation approach, where the server sorts clients’ models based on their local training losses before swapping them all consecutively according to the best and worst-performing ones. Our proposed approach is applied to a smart building dataset and compared with two other FL approaches from the literature. Our simulation results demonstrate improved forecasting precision for each client and faster convergence. Moreover, we optimized the global model’s evaluation error scores and overall loss, reduced the communication rounds required for convergence, and ensured less bias and more fairness between clients during each training cycle.

Federated learning (FL) enables collaborative machine learning among multiple devices without sharing private data. However, FL systems are vulnerable to poisoning attacks where malicious participants send malicious model updates to compromise the global model's accuracy. To enhance malicious model detection, we propose an EMDG-FL approach that optimizes the threshold used to identify attacks through a Genetic Algorithm (GA). The threshold indicates the degree of divergence between benign and malicious model updates. A tightly tuned threshold improves detection efficiency by reducing false positives and negatives. Our approach also includes a comparison study evaluating EMDG-FL against other defenses from literature across metrics like Accuracy Rate (ACC), Attack Success Rate (ASR) and Loss Rate (LR). Simulation results using two datasets demonstrate that EMDG-FL outperforms prior works in detecting poisoning attacks in FL. The optimized threshold calculation enables more precise and efficient identification of malicious models.

Today's scientific and corporate communities are highly interested in Wireless Sensor Networks (WSNs) and the Internet of Things (IoT). This kind of network consists of sensors with low resources that gather information for various real-life applications (healthcare, industrial, security, etc.), with streaming data requiring online processing. However, since outliers may occur in sensors collected data, it is necessary to identify and classify them into errors and events using online outlier detection and classification techniques suitable for the WSNs real-life applications. In this paper, we propose a centralised method for online outlier detection and classification in WSN. Our approach can differentiate between errors caused by malfunctioning sensors and errors caused by events. We also consider the spatial-temporal connection between sensor data vectors and nearby sensor nodes. Our approach, titled O2DCA, for Online Outlier Detection and Classification Approach, combines the benefits of the Fixed Width Clustering (FWC) and the Inter-Cluster Distance (ICD) algorithms for clustering outlier detection, respectively. For classification, we use the Inverse Distance Weighting (IDW) method, which allows us to classify outliers into errors that will be discarded and relevant events for which a necessary decision must be taken. We show through simulation using both synthetic and real-world datasets that our novel online approach is suitable for working with real-life applications where the Detection Rate (DR) performance metric stays stable and better than the offline approach.

In today's time, Wireless Sensor Networks (WSNs) and Internet of things (IoTs) have attracted a lot of interest from scientific and businesses communities. They are made up of limited-resource sensors that collect data for various applications (medical, manufacturing, militarily, etc.). However, data collected by sensors are susceptible to have outliers, which need to be detected and classified into errors and events using outlier detection and classification methods. In this paper, we propose a centralized outlier detection and classification approach for WSN. Our solution can distinguish between errors due to a faulty sensor and those due to an event. We also consider the spatial-temporal correlation between sensors' data values and neighbouring sensor nodes. Our approach, titled O2DNC for OPTICS-Based Outlier Detection with Newton Classification, combines the benefits of the OPTICS algorithm with a new method for outlier detection based on computing the variance and the average of the reachability distances. Furthermore, O2DNC uses a new approach based on the Newton interpolation and the K-Nearest Neighbours (KNN) algorithms to classify the outliers. For evaluation, we conduct a comparison study between our approach and two works from the literature and thus for the multivariate data case. Simulation results with both synthetic and real-life datasets show that the O2DNC outperforms the studied techniques in terms of several metrics like Detection Rate (DR), False Alarm Rate (FAR) and Receiver Operating Characteristic (ROC) curve.

Wireless Sensor Network (WSN) is one of the main components of the Internet of things (IoT) for gathering information and monitoring the environment in a variety of applications (medical, agricultural, manufacturing, militarily, etc.). However, data collected and transferred from sensors to the base station are susceptible to have outliers. These outliers can occur due to sensor nodes itself or to the harsh environment where they are deployed. Thus, it is necessary for the WSN to be able to detect the outliers and take actions in order to ensure network quality of service (in terms of reliability, latency, etc.) and to avoid further degradation of the application efficiency. In this paper, we propose a distributed outlier detection and classification algorithm for WSN. Our approach is capable of distinguish between an error due to a faulty sensor and an error due to an interesting event. We take into consideration the spatial-temporal correlation between sensors' data values and between neighbouring sensor nodes. Simulations with both synthetic and real datasets showed that our proposed approach outperforms other techniques by obtaining high Detection Rate (DR) and low False Alarm Rate (FAR).