Mid Sweden University

Empowered by deep learning, semantic communication marks a paradigm shift from transmitting raw data to conveying task-relevant meaning, enabling more efficient and intelligent wireless systems. In this study, we explore a deep learning-based task-oriented communication framework that jointly considers classification performance, computational latency, and communication cost. We evaluate ResNets-based models on the CIFAR-10 and CIFAR-100 datasets to simulate real-world classification tasks in wireless environments. We partition the model at various points to simulate split inference across a wireless channel. By varying the split location and the size of the transmitted semantic feature vector, we systematically analyze the trade-offs between task accuracy and resource efficiency. Experimental results show that, with appropriate model partitioning and semantic feature compression, the system can retain over 85 % of baseline accuracy while significantly reducing both computational load and communication overhead.

Federated learning (FL) enables privacy-preserving training on decentralized data but faces challenges on resource-constrained IoT devices due to high energy and communication costs. We address these by evaluating optimization strategies on a physical testbed of heterogeneous IoT devices. Our holistic approach, combining Top-K gradient compression with adaptive early stopping, reduces network bandwidth by 75% and lowers computational load without compromising accuracy. A compute-aware partitioning variant further rebalances workloads to minimize idle time. We identify the ”straggler” problem as a critical bottleneck, with faster devices idling for over 75% of the time. Finally, we provide a rigorous metrological assessment, quantifying Type A and Type B uncertainties to validate our findings. These findings highlight the need for system-aware optimizations for sustainable FL in IoT ecosystems.

Post-Quantum Cryptography is approaching deployment in TLS, raising important questions for the IoT, where large numbers of long-lived and resource-constrained edge devices must remain secure over decades. Many IoT systems rely on low-cost, Linux-capable edge nodes with limited compute, memory, and power resources, making post-quantum TLS a systems-level concern rather than a purely cryptographic one. This paper benchmarks representative classical and post-quantum TLS primitives on Raspberry Pi 1, 3, 4, used as proxies for constrained IoT edge devices bridging microcontrollers and cloud infrastructure. Performance is evaluated in terms of latency, hardware counters, cache behaviour, and peak resident memory across relevant message sizes. Results show that lightweight symmetric primitives such as Ascon-128 consistently outperform traditional TLS AEADs, while ML-KEM-512 and hybrid signature configurations provide the most balanced trade-offs among post-quantum public-key schemes. A PQC-enabled TLS-like handshake introduces modest overhead on modern IoT-class hardware but significantly higher cost on older platforms, highlighting the impact of device heterogeneity in deployed IoT fleets. Overall, the results demonstrate that post-quantum TLS is feasible on contemporary IoT edge hardware when algorithms are carefully selected, providing practical guidance for quantum-safe IoT deployments.

The Matrix Profile is a widely used framework for motif discovery and anomaly detection in time series due to its simplicity, efficiency, and parameter-free design. In many IoT and industrial monitoring applications, sensor nodes generate multivariate time series in which multiple measured quantities are jointly observed and often exhibit heterogeneous variance and statistical dependence. This work explores a geometry-aware extension of the Matrix Profile for multivariate IoT sensor measurements by incorporating Mahalanobis distance as the underlying similarity measure. By integrating a covariance-derived metric into the distance computation, the proposed formulation enables subsequence comparisons that account for variance anisotropy and cross-dimensional structure while preserving the fundamental semantics of the Matrix Profile framework. Experiments on real-world multivariate dataset illustrate the potential of geometry-aware similarity for motif discovery and anomaly detection beyond independent dimensions in multivariate IoT time series.

This study presents a large-scale field evaluation of "Kilea," a Retrieval-Augmented Generation AI tutor designed to support students with grounded explanations based on official study materials. While generative AI offers new possibilities for personalized education, its measurable impact on academic performance remains under-explored. Using a quasi-experimental difference-in-differences design, we analyzed over 200,000 graded evaluations from over 12,000 students to assess the system's efficacy. The results indicate high user acceptance (Net Promoter Score of +39) and verify a clear dose-response relationship: while casual usage yielded negligible benefits, high-intensity engagement correlated with statistically significant grade improvements. However, the overall net effect on learning outcomes across the entire cohort was modest (~0.04 grade points). These findings suggest that content-level personalization alone is insufficient to drive substantial performance shifts at scale. Regarding engineering requirements of these types of solutions, this study identifies the need to extend AI tutors with psychologically informed diagnostic models and 'ethics-by-design' principles. We conclude that future AI tutors must evolve beyond information retrieval, and move more towards data-driven, psychologically informed learner support that actively fosters self-regulation and motivation.

In today’s complex IT environments, understanding interconnected systems poses significant challenges. The nature of these systems introduces risks, making automatic observability and problem prevention critical. Machine learning and artificial intelligence enable promising solutions for managing this complexity. This paper addresses the challenge of detecting log anomalies within unstructured, continually changing logs that are difficult for traditional methods to parse. We propose a novel approach called LogDEC that utilizes joint clustering and dimensionality reduction together with sequential anomaly detection. Our method employs a sequential auto-encoder to detect anomalies through the reconstruction error of log feature sequences. Anomalies are scored by a threshold on the reconstruction error. LogDEC is trained unsupervised and can effectively identify anomalies, as demonstrated on a dataset collected at the Swedish Social Insurance Agency, as well as with three public log datasets. The model successfully and efficiently detected anomalies arising from manual interference, highlighting its applicability in real-world scenarios. 

Generative AI is rapidly reshaping higher education, and in this article we investigate how it can enhance learning rather than erode. In this paper we report findings from five discipline-specific pilots (computer engineering, law, mathematics, psychology, and teacher education) at Mid Sweden University guided by the university’s policy for ethical AI use. Each pilot explored AI as a student-centred learning tool and education tool, but from different perspectives and conditions based on subject. Across the pilots, we have found that these new tools can scaffold reflection, self-regulation, and engagement when integrated into course tasks. As well as need for clarity towards the students are important, to show when and how they can use these tools. We conclude that AI, when framed by robust pedagogy and clarity, can act as a catalyst for deeper learning and improved throughput. Future work will examine long-term retention and pathways to scaling these approaches university-wide. 

Integrating cyber-physical systems and the Internet of Things into industrial operations has significantly improved automation, efficiency, and data-driven decision making. However, these advances have also made industrial environments more vulnerable to cybersecurity risks. Our previous work explored lightweight deep learning models for real-time intrusion detection systems on edge devices, yet these models often operate as black boxes, limiting their trustworthiness. This issue is especially critical in the European Union, where the AI Act mandates transparency, accountability, and human oversight for AI solutions to be interpretable. In this paper, we integrate explainable AI solutions into lightweight real-time intrusion detection systems on edge devices to enhance the transparency and interpretability of black-box models. The study demonstrates that integrating SHapley Additive exPlanations significantly enhances the interpretability of intrusion detection systems, providing more transparent insights into model decision-making processes while maintaining accuracy and computational efficiency. This work contributes to the development of more secure and trustworthy industrial ecosystems by improving the effectiveness and reliability of intrusion detection.

The fast expansion of AI within distributed computing environments emphasizes questions about trustworthiness, particularly in contexts involving sensitive data and resource- constrained edge devices. To address this, we implement and evaluate a lightweight federated learning intrusion detection system in a realistic smart home scenario that combines TensorFlow Lite inference on edge devices, MQTT-based publishing, and co- ordinated training via the Flower framework. By operationalizing a previously proposed taxonomy and ontology of trustworthy AI in distributed systems, the implementation in this paper demonstrates how key trust dimensions, such as data integrity, model reliability, and process transparency, can be realized in edge environments. Our implementation utilizes local inference with TensorFlow Lite on IoT devices and coordinated federated evaluation via the Flower framework. We also introduce a trust score to quantify how the implementation aligns with the trust principles. The results indicate that the trust mechanisms are maintained without compromising accuracy or loss, contributing to practical insights into the application of theoretical trust frameworks within distributed AI systems.

The increasing need for robustness, reliability, and determinism in wireless networks for industrial and missioncritical applications is the driver for the growth of new innovative methods. The study presented in this work makes use of machine learning techniques to predict channel quality in a Wi-Fi network in terms of the frame delivery ratio. Predictions can be used proactively to adjust communication parameters at runtime and optimize network operations for industrial applications. Methods including convolutional neural networks and long short-term memory were analyzed on datasets acquired from a real Wi-Fi setup across multiple channels. The models were compared in terms of prediction accuracy and computational complexity. Results show that the frame delivery ratio can be reliably predicted, and convolutional neural networks, although slightly less effective than other models, are more efficient in terms of CPU usage and memory consumption. This enhances the model's usability on embedded and industrial systems.