📚 Recent Research Publications

Edge-cloud continuum has become a critical paradigm for supporting latency-sensitive applications that depend on efficient coordination among distributed resources. However, unpredictable hardware or software faults can disrupt services, leading to downtime degradation, performance degradation, and reduced user satisfaction. Enhancing the resilience of edge-cloud computing systems against unpredictable failures remains a critical and challenging issue. This paper introduces FT-PSOGA, a fault-tolerant edge-cloud service placement strategy that integrates Particle Swarm Optimisation (PSO) with a Genetic Algorithm (GA) to enable rapid and efficient redeployment of failed service components. By combining the fast convergence of PSO alongside the global search capabilities of the GA, the proposed method aims to minimise service recovery time while maintaining post-recovery performance in terms of service response time and reliability. FT-PSOGA is evaluated across diverse failure scenarios and system scales and benchmarked against several single- and multi-objective algorithms. The experimental results show that FT-PSOGA can rapidly restore the system in the event of failures and maintain post-recovery performance across different system scales and failure ratios (for example, recovering 60 failed components out of 600 in under six seconds). It minimises the service recovery time without compromising other performance metrics, showing FT-PSOGA as a robust and effective solution for resilient service placement in failure-prone edge-cloud environments.

Sequential decisions often involve trade-offs between immediate rewards and future risks, particularly in approach-avoidance contexts. While such behavior is commonly framed in terms of optimization principles, it remains unclear how decision processes adapt across contexts. We developed a sequential foraging task in which participants made binary choices under probabilistic reward and predation risk (threat). The design dissociated reward probability from threat and enabled formal comparison of decision-relevant features (e.g., reward and threat probabilities), a multi-feature policy, and a mathematically optimal policy derived from a fully observable Markov Decision Process. Crucially, the task included two implicitly signaled conditions - approach and avoidance - that differed in how reward and threat information jointly shaped the optimal policy. In both conditions, higher reward probabilities were associated with higher predation risk. However, the balance between these competing factors differed across environments, creating contexts in which the optimal policy either favored approaching or avoiding the higher-threat option. Behavior was analyzed using hierarchical Bayesian models capturing both individual decision features and optimal action values, as well as their modulation by task context.Avoidance contexts selectively altered the weighting of decision features, with reduced reliance on reward probability. In addition, choices showed evidence for increased alignment with optimal state-action values, even after accounting for heuristic features within a shared model. This indicates that behavior more strongly reflected the value structure of the environment under avoidance, consistent with enhanced integration of decision-relevant features.These findings suggest that adaptive behavior in sequential decision-making arises from context-dependent feature reweighting alongside increased alignment with integrated value signals. This provides a parsimonious account of how humans adjust decision-making under threat, without requiring assumptions about discrete strategy shifts.

The use of Machine Learning (ML) is rapidly expanding across diverse scientific and engineering domains. ML offers a powerful advantage over traditional modeling approaches for predictive modeling and analysis of variables of interest. This makes it particularly useful for developing advanced methods in analytical chemistry and residential energy systems, including forecasting (such as predicting hot water demand or chromatographic peak behavior), data quality assessment (such as detecting sensor drift or anomalous consumption patterns), and fault detection (such as identifying heat pump malfunctions or degraded separation performance). While traditional modeling approaches struggle to fully exploit complex, high-dimensional features, existing ML studies in the target domains often rely on limited datasets and lack automated and adaptive frameworks capable of handling the scale, variability, and non-stationary data generated in real operational settings. The availability of large amounts of data in this digital era offers unprecedented opportunities for analysis. However, the successful application of ML depends critically on the quality, preparation, and robustness of ML models and their underlying data. In industrial systems, these requirements are shaped by several interacting factors, particularly feature representation, model robustness, and adaptation under changing conditions. The challenges addressed in this research include data quality management, model selection, robustness assessment, and adaptation under changing real-world conditions. The main contributions of this research are: (i) a semi-automatic data preparation workflow with domain-specific feature engineering for large-scale oligonucleotide chromatography datasets; (ii) an unsupervised quality-centric evaluation framework that automatically clusters input data by quality level without requiring labeled annotations; (iii) the FIUL-Data fault injection framework, which quantifies the resilience boundaries of ML models under controlled data degradation; and (iv) a composite adaptive framework that integrates predictive ML with anomaly detection to enable demand-driven heat pump management in residential energy systems. Together, these contributions demonstrate that reliable industrial ML is achieved not by increasing model complexity, but through systematic data-centric practices including structured data preparation, quality-aware pipelines, robustness testing, and adaptive learning, applied across two complementary industrial domains.

5G networks are increasingly exposed to Distributed Denial of Service (DDoS) attacks launched by mobile botnets exploiting user equipment (UE). Existing studies have treated malware propagation and DDoS impact separately, with no prior simulation work having modeled how Device-to-Device (D2D) malware spreads and translates into network-wide DDoS stress. This paper introduces DDoSimu5G, an extension of Simu5G [33] integrated with the ONE simulator [24], which enables, for the first time, the combined modeling of D2D malware propagation, UE mobility, and botnet-driven traffic loads on 5G infrastructure. The framework is aligned with 3GPP Proximity based Services (ProSe) in the 5G System (5GS) TS 23.304 [1] specifications, supports configurable attack scenarios, and generates diverse datasets, including infection logs, mobility traces, and PCAP traffic. By linking propagation dynamics with DDoS effects, DDoSimu5G provides the research community with a reproducible open-source tool for studying and mitigating emerging D2D-driven threats in 5G networks.

We present DSIPTS-P, a Kubernetes-native platform for interactive, distributed spatiotemporal forecasting across edge-cloud infrastructures. DSIPTS-P enables split learning and federated-style execution while automating the full ML lifecycle, including deployment, training, monitoring, and recovery. Built on the DSIPTS stack, the platform provides GPU-and edge-aware scheduling of training and inference pods, declarative YAML-based or GUI-based configuration, and zero-touch deployment with integrated telemetry through Prometheus, Grafana, Kepler, and AIM. In the live demo, attendees actively configure forecasting models, split learning partition points, and hardware placement of training and inference pods, then trigger failures and network impairments on the fly. Real-time dashboards immediately reveal how these choices impact forecasting accuracy, convergence behavior, communication overhead, latency, and energy consumption across heterogeneous edge and cloud resources.

Large-scale agile (LSA) frameworks are increasingly adopted by financial institutions to balance rapid digital innovation with complex regulatory requirements. However, the systemic impact of these frameworks on developer cognitive capacity, psychological safety, and long-term socio-technical sustainability remains under-researched. This paper presents exploratory case study of a Swedish financial services organization navigating a Scaled Agile Framework (SAFe) transformation. We conducted fifteen semi-structured interviews across multiple organizational roles and analyzed the data using reflexive thematic analysis. Our analysis identify six recurring socio-technical tensions shaping developer experience and team functioning. Rather than representing isolated challenges, these tensions form an interacting system. We synthesize these findings to explain how coordination intensity, governance ambiguity, and delivery pressure recursively reinforce cognitive overload, defensive team behavior, and socio-technical debt accumulation. We argue that developers well-being is a leading indicator of socio-technical sustainability, not a peripheral human resources concern.

Network diversity, high reliability, intelligence, and the Internet of Things (IoT) are at the forefront of the distributed Edge-native computing vision, which aims to provide orchestration, better resource utilization, and self-optimizing networks. The integration of emerging disruptive technologies, namely AI, network autonomy, and a hybrid cloud computing architecture, with edge-native computing, propels the rapid development of an upgraded service-oriented architecture at the network’s core/edge level. This acceleration is intended to support on-demand microservices, specifically focusing on visibility services essential for intelligent network management. This article explores the significance of AI-enabled ENC, including its design requirements and projected benefits. To address the challenges of managing and optimizing complex, heterogeneous distributed networks, we present a use case study and experimental evaluation based on the design of an edge intelligence framework. This framework leverages resources of a multisite cloud/edge-native microservices testbed to enhance network visibility and automation, enabling the deployment and orchestration of AI-powered microservices at the network edge. By applying statistical, machine learning, and deep learning-based analyses to this diverse dataset, we compare the accuracy of these forecasting models and demonstrate the effectiveness of our approach in accurately predicting network traffic patterns, even under varying network conditions. This accuracy is validated through low root-mean-square error and high coefficient of determination values. Our results highlight the potential of AI-enabled edge-native computing to revolutionize network management and optimization, enabling timely detection of performance degradation and proactive mitigation strategies.

The edge-cloud paradigm improves service delivery by orchestrating resources across edge nodes and cloud data centres. These environments consist of heterogeneous, interconnected computing nodes that cooperate to deliver continuous services. However, their scale and complexity increase vulnerability to failures from hardware malfunctions, software defects, and dynamic operating conditions. These failures can disrupt system configurations and service execution, leading to reduced reliability, performance degradation, and violations of service-level objectives. Ensuring service execution requires adaptive service placement strategies across edge-cloud resources. This study introduces a fault-tolerant service placement approach (Enhanced Evolution Strategy for Collaborative Neural Decision-making, EES-CND) for edge-cloud environments. The method employs collaborative decision-making, wherein multiple lightweight neural networks jointly infer redeployment strategies during failure events. To address the system dynamics and mitigate performance drift, adaptive models are updated online using an enhanced evolution strategy. Extensive simulations show that EES-CND effectively handles performance drift and significantly outperforms existing methods in service recovery time, response time, and reliability, achieving a 44.8% reduction in fault-tolerance cost compared to standalone models.

In modern video streaming systems, clients typically make independent bitrate decisions without full knowledge of network conditions, often leading to inefficient resource usage and unstable quality. Network-assisted adaptive streaming is rapidly gaining importance in Network and Service Management (NSM), as operators strive to deliver consistently high Quality of Experience (QoE) under dynamic traffic conditions. This paper introduces FROG, a novel, real-time Software-Defined Networking (SDN)-assisted framework designed to coordinate multi-server and multipath HTTP Adaptive Streaming (HAS) using standardized metrics carriage: Common Media Client Data (CMCD) and Common Media Server Data (CMSD). FROG employs an innovative two-stage optimization workflow in which an initial Linear Programming (LP) model rapidly determines feasible bandwidth bounds, server selection, and path capacities, thereby transforming the remaining optimization into a sequential layer-selection process for coordinated quality selection and flow allocation. This decomposition enables sub-second optimization at the scale of thousands of users, while buffer-aware client feedback is integrated to proactively prevent stalls and maintain system stability. Experiments on an emulated SDN testbed demonstrate that FROG achieves QoE comparable to that of the optimal MILP solution on tractable instances while virtually eliminating video stalls, and significantly reduces quality oscillations, outperforming state-of-the-art network-assisted approaches by up to 2.32× under playback-driven evaluation scenarios. Scalability experiments with up to 4,000 clients further demonstrate sub-second optimization runtimes, confirming the practicality of FROG for large-scale deployments.

The transition towards Industry 5.0 can leverage Kubernetes-driven computing infrastructure to support cyber-physical systems and IoT-based automation of containerized microservices. However, usability challenges of Kubernetes (e.g., fragmented workflows and syntax-sensitive configuration languages) hinder effective human-machine collaboration. To address these challenges, we present a web-based orchestration system prototype that streamlines containerized workloads management through guided workflows, multimodal features, and real-time monitoring. By reducing dependence on command-line expertise and providing a user-friendly interface, the prototyped system enhances the reliability and transparency of interactions between humans and complex computing platforms. We evaluated the prototyped system through a mixed-methods usability study including task analysis, the System Usability Scale (SUS), and interviews. Results demonstrate significant improvements in usability, with SUS score of 77.25 indicating "Good" overall usability. Overall, these findings show that usability-centered orchestration interfaces can strengthen human trust, improve collaboration, and accelerate the adoption of cloud-native platforms as key enablers of Industry 5.0.

3GPP Access Traffic Steering, Switching, and Splitting (ATSSS) enables traffic to be distributed across heterogeneous 3GPP and non-3GPP access networks to improve performance, reliability, and resilience. ATSSS can use multipath transport protocols such as Multipath QUIC (MP-QUIC), where packet scheduling plays a central role in determining latency and resource utilization for delay-sensitive applications. Many existing MP-QUIC scheduling policies rely on instantaneous path measurements or fixed rules rather than forecasts of future application demand. In multi-flow scenarios, such decisions can lead either to contention on the preferred low-latency path and transient latency inflation for priority traffic or to overly conservative use of available capacity. This paper proposes iScavenger, a predictive, machine-learning-based multi-flow scheduling policy for ATSSS environments. iScavenger employs a Long Short-Term Memory (LSTM) model to forecast near-future bandwidth demand for delay-sensitive Sticky traffic. Based on this prediction, background packets are admitted to the preferred low-latency path only when sufficient residual capacity is expected to remain; otherwise, they are steered to the alternative path. The policy is implemented within the Monty MP-QUIC framework and evaluated in a controlled Mininet testbed using traffic traces from the online game League of Legends, with fixed and variable path capacities and controlled jitter and packet loss. The results indicate that, under the evaluated conditions, iScavenger provides configurable operating points in the latency–utilization trade-off, limiting additional Sticky-flow RTT while achieving higher background throughput than conservative baseline policies. These findings highlight the potential of short-term traffic-demand prediction for proactive contention management in ATSSS-enabled multi-access networks.

Modern high-speed packet processing systems must enforce strict Quality-of-Service (QoS) guarantees on loss, latency, and throughput across many concurrent flows, typically using hierarchical QoS (HQoS) schedulers. In NFV environments, software schedulers provide fine-grained control, but maintaining QoS at line rate requires careful coordination between software and hardware. We show that uncoordinated per-core software schedulers can oversubscribe a shared NIC, causing packet drops in NIC transmit queues that violate HQoS guarantees even when multiple hardware priority queues are available. To address this, we propose Priority-Aware Backpressure (PAB), a host-NIC coordination mechanism that feeds NIC congestion signals back to the software scheduler to selectively throttle low-priority traffic while protecting high-priority flows. We implement PAB in the DPDK QoS framework and leverage the shaping hardware capabilities of a commodity NIC. We demonstrate that PAB preserves hierarchical QoS guarantees under sustained overload, while maintaining fairness, low latency for high-priority traffic and only modestly reducing throughput compared to uncoordinated schedulers. © 2026 IEEE.

The rights to rectification and erasure, as established under the General Data Protection Regulation (GDPR), are central to protecting individuals’ privacy. However, their effective enforcement in machine learning (ML) systems remains challenging. Existing work has largely addressed these rights from either a legal or a technical perspective in isolation and disregards the fact that models are produced in complex supply chains involving multiple actors across development, distribution, and deployment. This paper presents a holistic survey of challenges in implementing the rights to rectification and erasure in ML models. Drawing on academic literature and guidance from data protection authorities, we find that many GDPR requirements cannot yet be technically met in practice. Our findings further suggest that issues arising in ML supply chains are insufficiently addressed in research. To tackle this gap, we introduce the notion of models in the dark – derived models created further downstream in an ML chain without sufficient transparency or traceability – and analyse the urgent challenges posed by this phenomenon. By adopting an interdisciplinary perspective, this work contributes to bridging the gap between legal requirements and the technical implementation of data subject rights in ML, ultimately supporting the development of trustworthy artificial intelligence.

Context: Large-scale agile (LSA) is inherently characterized by socio-technical complexities (e.g., system dependencies, extensive cross-team coordination, and distributed organizational structures). Onboarding of newcomers is a critical challenge in LSA environments and it remains underexplored compared to small-scale agile settings. Objective: Our study aims to investigate the onboarding processes and challenges within LSA projects. Method: In this exploratory qualitative study, we conducted 41 semi-structured interviews across four Swedish software companies. Results: We identified onboarding issues that result in negative outcomes, such as cognitive overload, psychological safety constraints, and cross-team collaboration hurdles. These factors create steep learning curves and knowledge deficiencies that contribute to Conclusions: We provide actionable recommendations for practitioners to improve their onboarding processes. In particular, we recommend designing competency-based rather than time-based onboarding plans, expanding social integration through structured cross-team exposure, and addressing the gap between textbook agile and real-world LSA practices through explicit expectation setting.

As 5G networks adopt cloud-native architectures, operational complexity motivates automation of several parts of the network, especially fault management. While Large Language Models (LLMs) show promise as autonomous operators, current evaluations rely on static benchmarks that measure plausibility rather than functionality. We introduce OperAID, an open-source testbed that evaluates LLM agents through a closed-loop pipeline comprising fault injection, diagnosis, remediation, and executionbased verification on a live Open5GS deployment in Kubernetes. Across 900 experiments spanning three fault scenarios with five open-source LLMs, we find that: (i) tool access increases average performance from 10.9 % to 61.1 %; (ii) model size alone does not guarantee success even with tools; and (iii) without tools, models fail to diagnose faults requiring live system observation.