As 5G networks evolve toward disaggregated Open Radio Access Network (O-RAN) architectures, low-latency, edge-resident security mechanisms are critical. Conventional Extended Application (xApp)-based Intrusion Detection Systems (IDS) at the Central Unit (CU) create high computational overhead and latency due to their reliance on deep packet inspection and centralized analytics. This work proposes a lightweight, Distributed Application (dApp)-based IDS deployed at the Distributed Unit (DU) for near-real-time threat detection at the radio edge. Our approach leverages radio telemetry features from the E2 interface to identify network attacks without inspecting upper-layer packet data. Using a live 5G O-RAN testbed, we constructed synchronized datasets from the DU (lower-layer) and CU (upper-layer) under various attack scenarios to evaluate several Machine Learning (ML) models. The results demonstrate that our dApp-based IDS achieves comparable accuracy to traditional xApp systems while significantly reducing CPU effort, memory usage, and execution time, underscoring the potential of dApps as scalable and effective security primitives in O-RAN.

The convergence of immersive technologies and next-generation communication networks offers new potential for advancing surgical training. This paper presents the Magos Bakri Balloon Placement Training (MBBPT) system, a Virtual Reality (VR) simulation platform integrated with haptic feedback and 5th Generation (5G) Open Radio Access Network (O-RAN) connectivity. The system enables realistic and interactive medical training, leveraging submillimeter-precision hand tracking and kinesthetic feedback from Magos gloves. To evaluate the feasibility and performance of network-assisted VR training, MBBPT was tested across a disaggregated O-RAN-based private 5G testbed at University College Dublin (UCD), a standalone private 5G testbed at Patras, and a cross-site setup linking both. The setup assessed Key Performance Indicators (KPIs) and Key Value Indicators (KVIs) to show the system supports responsive, multiuser VR interactions across geographically distributed sites, with consistent performance. These findings highlight the role of 5G O-RAN as an enabler for scalable, collaborative, high-fidelity immersive medical education.

We demonstrate SPATIAL, a proof-of-concept system that augments modern applications with capabilities to analyze trustworthy properties of AI models. The practical analysis of trustworthy properties is key to guaranteeing the safety of users and overall society when interacting with AI -driven applications. SPATIAL implements AI dashboards to introduce human-in-the-loop capabilities for the construction of AI models. SPATIAL allows different stakeholders to obtain quantifiable insights that characterize the decision making process of AI. This information can then be used by the stakeholders to comprehend possible issues that influence the performance of AI models, such that the issues can be resolved by human operators. Through rigorous benchmarks and experiments in a real-world industrial application, we demonstrate that SPATIAL can easily augment modern applications with metrics to gauge and monitor trustworthiness. However, this, in turn, increases the complexity of developing and maintaining the systems implementing AI. Our work paves the way towards augmenting modern applications with trustworthy AI mechanisms and human oversight approaches.

Despite its enormous economical and societal impact, lack of human-perceived control and safety is re-defining the design and development of emerging AI-based technologies. New regulatory requirements mandate increased human control and oversight of AI, transforming the development practices and responsibilities of individuals interacting with AI. In this paper, we present the SPATIAL architecture, a system that augments modern applications with capabilities to gauge and monitor trustworthy properties of AI inference capabilities. To design SPATIAL, we first explore the evolution of modern system architectures and how AI components and pipelines are integrated. With this information, we then develop a proof-of- concept architecture that analyzes AI models in a human-in-the- loop manner. SPATIAL provides an AI dashboard for allowing individuals interacting with applications to obtain quantifiable insights about the AI decision process. This information is then used by human operators to comprehend possible issues that influence the performance of AI models and adjust or counter them. Through rigorous benchmarks and experiments in real- world industrial applications, we demonstrate that SPATIAL can easily augment modern applications with metrics to gauge and monitor trustworthiness, however, this in turn increases the complexity of developing and maintaining systems implementing AI. Our work highlights lessons learned and experiences from augmenting modern applications with mechanisms that support regulatory compliance of AI. In addition, we also present a road map of on-going challenges that require attention to achieve robust trustworthy analysis of AI and greater engagement of human oversight.

With the rapid progression of communication and localisation of big data over billions of devices, distributed Machine Learning (ML) techniques are emerging to cater for the development of Artificial Intelligence (AI)-based services in a distributed manner. Federated Learning (FL) is such an innovative approach to achieve a privacy-preserved AI that facilitates ML model sharing and aggregation while keeping the participants’ data at the original source. However, recent research has investigated threats from poisoning attacks in FL. Several robust algorithms based on techniques such as similarity metrics or anomaly filtering are proposed as solutions. Yet, these approaches do not focus on investigating the intentions of the attackers or providing justifications and evidence for suspecting the behaviour of clients who are considered poisoners. Therefore, we propose SHERPA, a robust algorithm that uses Shapley Additive Explanations (SHAP) to identify potential poisoners in an FL system. Based on this, we develop a novel algorithm to differentiate poisoners via feature attribution clustering. We launch data poisoning attacks for different scenarios on multiple datasets and showcase our solution to mitigate the attacks. Furthermore, we show that privacy-targeted poisoning attacks can be mitigated with our approach. Accompanying the Explainable AI (XAI) technique for defence, our study reveals the potential for post-hoc feature attributions in countering data poisoning attacks with better explainability and improved justification in eliminating potentially malicious clients in the aggregation process.

Federated Learning (FL) is an emerging privacy-preserved distributed Machine Learning (ML) technique where multiple clients can contribute to training an ML model without sharing private data. Even though FL offers a certain level of privacy by design, recent works show that FL is vulnerable to numerous privacy attacks. One of the key features of FL is the continuous training of FL models over many cycles through time. Observing changes in FL models over time can lead to inferring information on changes to private and sensitive data used in the FL process. However, this potential leakage of private information is not yet investigated significantly. Therefore, this paper introduces a new form of inference-based privacy attacks called FL Time Inference Attacks (FL-TIA). These attacks can reveal private time-related properties such as the presence or absence of a sensitive feature over time and if it is periodical. We consider two forms of such FL-TIA: i.e. identifying changes in membership of target data records over training rounds and detecting significant events in clients over time by observing differences in FL models. We use the network Intrusion Detection System (IDS) as a use case to demonstrate the impact of our attack. We propose a continuous updating attack model method for membership variation detection by sustaining the accuracy of the attack. Furthermore, we provide an efficient detection method that can identify model changes using cosine similarity metric and one-shot mapping on shadow model training.