Microcontroller-based neural network inference faces significant RAM constraints, hindering performance and deployment. One of the main constraints is the peak memory usage, which is essential for conducting deep learning inferences with low latency. To address this, previous researchers have developed peak memory estimators to estimate the Peak Memory Usage, which could be used by inference-time optimization techniques like pruning to tackle the RAM constraint. But many of the peak memory estimators used by current state-of-the art frameworks like µNAS and TinyEngine produce underestimation or overestimation, reducing the reliability of model decisions made under RAM constraints. Underestimation often arises from failing to account for all components contributing to peak memory usage, while overestimation can occur when extra memory overheads irrelevant in MCU-specific inference scenarios are incorrectly included. In this paper, we propose our peak memory estimator, which estimates the peak memory usage of deep learning inference at the operator level and can accurately estimate the peak memory usage of a deep learning model during inference. The experiments show that our method achieves more accurate peak memory predictions across multiple MCU platforms and can be effectively integrated with pruning strategies to produce better model compression that satisfies both memory and accuracy constraints. Specifically, on a benchmark of 150 models, our method achieved an average estimation error margin of only 0.9%, significantly outperforming µNAS, which exhibited an average error margin of 61.7%.

The sixth generation (6G) of mobile networks promises transformative capabilities in terms of, amount others, higher data rates, lower latency and ubiquitous coverage, but achieving these goals sustainably poses significant challenges. A promising solution lies in Cell-Free massive Multiple-Input Multiple-Output (CF-mMIMO) networks, where a dense deployment of geographically distributed Access Points (APs), coordinated by one or more Central Processing Units (CPUs), cooperatively serve User Equipments (UEs). While CF-mMIMO networks offer improved spectral efficiency and more spatially uniform service, their dense deployments can result in high energy consumption. As MNOs typically design their networks such that Quality of Service (QoS) targets are met during peak hours, the network is left significantly over-dimensioned during off-peak hours. This thesis addresses this inefficiency by proposing a low-complexity, heuristic Sleep Mode Management (SMM) algorithm that reduces energy usage by switching off unneeded APs while maintaining acceptable QoS and without compromising coverage.

The proposed SMM algorithm supports multiple AP power states: active, light sleep and deep sleep. It incorporates realistic transition times between these states. Relying solely on practically available information such as long-term Channel State Information (CSI) and previously achieved data rates, the algorithm dynamically decides which APs can be temporarily put into a sleep mode. Importantly, it ensures population coverage is maintained and it preserves UE QoS based on a 10th UE throughput percentile target.

The proposed SMM algorithm is evaluated using a system-level simulator that models a realistic scenario based on the city center of Amsterdam, including lamppost-based AP deployments and a realistic basis for the spatial traffic distribution and daily traffic fluctuations. Simulation results demonstrate that the proposed SMM algorithm reduces the daily energy consumption of a CF-mMIMO network by up to 17.11\% with the best overall configuration. If the parameters of the SMM algorithm are allowed to be adaptively tuned to the traffic load, the daily energy consumption can be reduced by up to 21.54%. This thesis not only contributes a novel SMM algorithm but also provides guidance for AP deployment strategies in 6G CF-mMIMO networks, determining that a higher number of lower-antenna-count APs can yield better QoS than fewer, more higher-antenna-count APs at the cost of energy efficiency.

The rise of Low-Earth-Orbit (LEO) satellite networks, such as Starlink, has transformed global connectivity, enabling high-speed internet access in previously underserved regions. However, existing research lacks a unified framework to evaluate and compare the performance of LEO ISPs against terrestrial alternatives using heterogeneous measurement datasets. In this work, we present a methodology for harmonizing and standardizing passive internet measurements from M-Lab’s NDT7 and Cloudflare’s AIM datasets, implemented in the form of the Global Telemetry System. These sources are integrated through schema unification, filtering, and normalization to produce a reproducible and geographically comprehensive telemetry dataset. We introduce a server-based filtering approach to mitigate geographic and routing biases, and we evaluate aggregation methods to align measurement distributions across datasets. Our results demonstrate the dataset integration methodology preserves key distributional properties, enabling fair and statistically consistent merging of measurements from the two sources. This work represents a first step toward a scalable and extensible telemetry infrastructure for assessing next-generation global internet services.

Low Earth Orbit (LEO) satellite constellations, particularly SpaceX’s Starlink, have quickly gained popularity and have become a viable alternative to traditional terrestrial Internet Service Providers (ISPs) in recent years. However, due to their novelty and unique architecture, research into their performance is limited, especially one comparing LEO and terrestrial internet. This paper will demonstrate how user-initiated active measurements can be used to both gather new data about LEO internet and assess and compare the performance of individual networks. First, performance metrics that reflect typical internet usage scenarios, such as web browsing and video streaming, are chosen. Next, a test suite is developed to collect data about one’s network. It also augments this data with location information to aid in later comparison. This data can then be used for comparisons between individual networks as well as in further research. The last step integrates the suit developed into a web based platform that aims to provide a wide variety of information about Starlink’s performance, architecture and allow users to gauge the potential benefits of transitioning from terrestrial to LEO internet could provide them.

We present a machine learning framework aimed at forecasting Starlink (LEO satellite) network performance at fine spatiotemporal resolution. Our approach combines MLab crowdsourced measurements, weather and forecast features, and dynamic satellite density to predict packet loss, jitter, latency, and throughput. We introduce a composite Weather Index and real-time satellite density per location, and train robust ensemble models with anomaly filtering and median aggregation. Our best models achieve good predictive results with less than 17 ms for latency, and 35 Mbps for throughput. Latency is reliably predictable with meteorological and satellite context, while packet loss and jitter remain challenging. Predictions are limited to periods close to the training data and our results establish a reproducible baseline for short-term, weather-aware Starlink network forecasting.

Recommender systems are widely used in modern lives and contribute to many industries. Therefore, methods to evaluate and improve them are important. Nowadays, much research has been done to improve the system aspects such as algorithms. However, user experiences are not only affected by the systems but heavily rely on the context when using the systems. Therefore, the research on user aspects to understand their experiences is as important. This study contributes an approach that uses collaborative reflection to find insights into users' experiences with recommender systems. Using this approach, this study presents the influences of context on user experiences with recommender systems. This study investigates the importance of situational and personal contexts like mood, time, and location in shaping user satisfaction with recommendations. The research adopts a method based on collaborative reflection, where participants engage in tasks using their YouTube watch history, paired with another individual for real-time discussion. By analyzing contextual influences and the values users wish to achieve, the study identifies key patterns in user behavior and insights into personal preferences. This research not only contributes to the evaluation of recommender systems but also highlights the need for systems to align with both the goals of users and broader societal values. The usability of the proposed method was tested to be successful in crowdsourcing, yielding practical implications for future evaluation and improvements of recommender systems.

Federated Learning (FL) is a distributed machine learning approach that enhances data privacy by training models across multiple devices or servers without centralizing raw data. Traditional FL frameworks, which rely on synchronous updates and homogeneous resources, face significant performance challenges in real-world deployments. These challenges include handling heterogeneous data distributions (non-IID) and varying computational resources across clients, leading to performance degradation. To address these issues, asynchronous FL frameworks have been proposed, but they introduce new complexities such as communication latency and workload management across geo-distributed servers.

This paper focuses on the impact of network latency and non-IID data distributions on asynchronous multi-server FL systems. We propose and evaluate three methods to mitigate the adverse effects of data heterogeneity on model accuracy: (1) transferring clients between servers; (2) sharing clients among servers so that they alternatively train their models; and (3) sharing clients among servers with model averaging. Our contributions include a reproducible experimental framework for multi-server FL, strategies for optimizing client-server interactions, and an analysis of the effectiveness of these strategies in reducing the impact of non-IID data distributions. Experimental results demonstrate that our methods can reduce training time by up to 85.67%, decrease the number of updates required by 85.82%, and improve accuracy by 4.815% in heterogeneous environments, compared to Spyker, a state-of-the-art multi-server FL algorithm.

Visible Light Communication (VLC) leverages the visible light spectrum to establish wireless communication, offering advantages such as broader bandwidth, and reduced energy consumption compared to traditional radio frequency methods. VLC offers two main approaches: passive and active. Passive VLC takes advantage of sunlight, which is pervasive and highly power-efficient. However, its reliability can be affected by weather conditions and the absence of sunlight at night. On the other hand, active VLC which uses artificial light sources like LEDs, provides more consistent performance but is not power-efficient when sunlight is available. For example, during the day when ample sunlight could be used for passive VLC, turning on a light bulb for active VLC is unnecessary and wasteful.

This thesis tackles these challenges by combining the best of active and passive systems to create an even more power-efficient and reliable system. It addresses two key problems in passive VLC: reducing the power consumption of passive VLC transmitters and enhancing the reliability of passive VLC links through a hybrid system. By replacing the FPGA-based controller with a low-power microcontroller, the power consumption of the Digital Micro-mirror Device used for sunlight modulation was significantly reduced from 1.3W to 36.85mW, while achieving a data rate of 25 kbps with a bit error rate (BER) of less than 1% at a distance of 25 cm. Its maximum range was determined to be 75 cm at 10 kbps. Additionally, integrating an LED component into the passive VLC communication link improved reliability in varying ambient light conditions. The hybrid system demonstrated enhanced performance in low ambient light scenarios, ensuring a BER below 1% regardless of ambient light conditions. In high ambient light scenarios, the LED can be dimmed or turned off, conserving power and making the system more efficient than a purely active VLC system. This thesis contributes to the advancement of energy-efficient and reliable VLC technologies, paving the way for their broader adoption.

Federated Learning is a machine learning paradigm where the computational load for training the model on the server is distributed amongst a pool of clients who only exchange model parameters with the server. Simulation environments try to accurately model all the intricacies of such a system. However, current simulators do not properly impose the concept of simulation time, leading to global model inaccuracies and difficulties of replicating reruns of the simulation, which is most prominent in the asynchronous scenarios. To this purpose, we propose a discrete-event simulator for the central server asynchronous case which timestamps all the events in the system prior to execution, reducing variability in client model updates on the server. We also introduce a log-structure used to keep states of the simulation, making client inspection possible based on time. We evaluate the proposed discrete-event simulator on the baseline simulator of Flower, reducing standard deviation amongst server model updates for 31.5% and improving accuracy with heterogeneous clients in the MNIST case for 3.3% on average.

Federated learning (FL) is a machine learning paradigm where private datasets are distributed among decentralized client devices and model updates are communicated and aggregated to train a shared global model. While providing privacy and scalability benefits, FL systems also face challenges such as client and data heterogeneity, where the training resources and different datasets are non-Independent and Identically Distributed (non-IID). When testing novel FL algorithms or configurations, simulators are used to create controlled environments without needing costly deployments. However, it is important to understand how representative FL simulators are of real-world deployments, and what steps can be taken to bridge the gap between theoretical results and practical implementations. In this paper, we investigate the effects of incorporating traces from a pseudo-real heterogeneous FL deployment in simulated environments. We compare four non-IID attributes, including batch sizes, local epochs, data volume, and data labels, to determine the most influential factors for reproducing the deployment results in simulations. We show that there is an inherent difference between deployments and simulations, despite incorporating identical non-IID conditions. Furthermore, we show that including non-IID data labels in simulations has the most significant impact on recreating the deployment outcome. We also demonstrate that incorporating the other mentioned factors has negligible impact, resulting in similar training performance compared to fully IID simulations. Our results are derived from a 20-client single-server synchronous FL configuration, and additional research is necessary to confirm our findings for larger-scale systems.

Federated Learning has gained prominence in recent years, in no small part due to its ability to train Machine Learning models with data from users' devices whilst keeping this data private. Decentralized Federated Learning (DFL) is a branch of Federated Learning (FL) that deals with clients directly communicating with each other as opposed to using a central server. Client mobility describes how users' devices move in the real world, and its effects on the learning performance of Hierarchical Federated Learning (HFL) systems have been found to be significant. However, the effects of client mobility on DFL systems have not been explored. In this work, we fill this research gap. First, we develop a model that can describe client mobility in a DFL system. Then, using synthetic datasets, we show that client mobility has a positive impact on learning performance, which we quantify. Moreover, we show that there is a disparity in learning performance between high-mobility and low-mobility clients when using a baseline model aggregation algorithm. To address this disparity, we propose a new mobility-aware model aggregation algorithm. Our experimental results on synthetic datasets show that our solution reduces the disparity in learning performance between high- and low-mobility clients in the scenarios where this disparity is greatest, with no appreciable downsides in global learning performance.

Federated Learning (FL) systems often suffer from high variability in the final model due to inconsistent training across distributed clients. This paper identifies the problem of high variance in models trained through FL and proposes a novel approach to mitigate this issue through scheduling simulations subject to precedence constraints. By effectively scheduling the execution of client tasks and parameter server updates, we aim to reduce the variance in the final aggregated model. Through a series of experiments, we demonstrate that our proposed scheduling method significantly reduces model variance, while not impacting the time of simulation drastically. Additionally, we propose 2 algorithms to solve the problem of scheduling under precedence constraints - Ant Colony Optimisation, and an Evolutionary Algorithm - to minimize the makespan of simulations.

Threshold signatures play a crucial role in the security of blockchain applications. An efficient threshold signature can be applied to enhance the security of wallets and transactions by enforcing multi-device-based authentication, as this requires adversaries to compromise more devices to recover the key. Additionally, threshold signatures can protect user privacy, for instance, by enabling anonymous transaction signing on behalf of a group of users sharing a blockchain wallet. This study conducts a comprehensive analysis of threshold signature schemes, identifying FROST as the premier choice due to its performance efficiency, improved energy consumption, and practical feasibility on mid-range IoT devices and smartphones as demonstrated through empirical testing. Furthermore, we introduce a protocol for integrating FROST with Hyperledger Fabric v3.0, aimed at enhancing IoT devices' ability to interact with blockchain networks through efficient transaction signing. Our experiments reveal that an IoT network of five devices, under optimal network conditions, can sign a transaction and commit it to the Hyperledger Fabric network within 3.2 seconds, leveraging a 2-second batch timeout configuration.

In the ever-evolving field of music technology, new solutions continue to emerge that enhance musical expression and creativity. This thesis introduces WaveTune, a novel lightweight hand gesture recognition system that enables real-time control of musical composition and performance through natural hand motions.

WaveTune utilizes Millimeter Wave radar technology to capture gesture data, combined with optimized deep learning techniques for real-time recognition. This provides an accessible and non-intrusive platform for gesture control that enhances privacy since no visual data is recorded. Users can dynamically select tracks and control musical parameters in real-time using expressive hand motions, integrating seamlessly with music software.

A key innovation of WaveTune is the development of an optimized gesture recognition model that achieves high accuracy for real-time music interaction while minimizing complexity. This is accomplished through novel optimizations to a state-of-the-art point cloud classification architecture, resulting in an efficient and tailored model for fluid musical control.

Furthermore, WaveTune promotes open-source collaboration by providing full access to code, configurations and datasets, inviting the community to build upon this system.

Drones that perform complex autonomous movements require a perfect estimate of their current position. However, internal measurement unit (IMU) errors introduce drift in this estimate, leading to significant discrepancies between the predicted and actual location. Various solutions have been proposed to calibrate the IMU, including methods involving cameras and humans in the loop. This thesis suggests implementing a previously developed technique that involves projecting a precise static light polarisation grid into a room. Although this pattern is invisible to the human eye it can be observed using a polariser and colour sensor combination. A drone equipped with such a sensor setup can recalibrate for IMU drift by utilising the perceived polarisation patterns as optical landmarks.

The system design is further developed by exploring the potential of visible light communication (VLC) as an alternative to traditional radio frequency (RF) links for drone control. By leveraging the existing infrastructure used for the projection of the polarisation grid, a VLC link is integrated into the system. With the addition this work strives to fuse polarisation-based localisation and VLC,
setting the first steps in creating a fully visible light-based drone platform.

To validate the system, a prototype is created that achieves real-time simultaneous localisation and communication on an embedded drone. This is accomplished through machine learning based classification, a drone motion model, an optimised polarisation pattern enabling fast localisation and a noise-resistant VLC link. Experiments show a median 2D tracking error of 10cm using only light-based methods and a VLC link range of up to 2.5 meters under various conditions.

Distributed Denial of Service (DDoS) leverages the power of multiple servers to disrupt the operations of a victim service. Due to the financial risks posed by downtimes on critical online infrastructure, DDoS is among the top threats in the cybersecurity landscape.

In this paper, we analyze the characteristics of previously launched DDoS attacks using collected network data. To extract the characteristics from a network trace file, we expand the DDoS Dissector tool with additional statistics representing the peak traffic strength and the sources of the attack. In addition, we implement an algorithm to parallelize the analysis of large-scale attacks when executed in memory-constrained environments. Our results show that the error difference in the statistics obtained when running the parallelized version and the original one is less than 0.5%.

Furthermore, we investigate several DDoS attacks by analyzing the contained attack vectors and their corresponding characteristics. Our software correctly detects the attack vectors, however, we remark that the output quality is impacted by the percentage of non-attack traffic. In particular, we provide an overview of the current state of the DDoS landscape seen from the point of view of a scrubbing service and study the effect of a Booter takedown on the frequency of DDoS attacks. Lastly, we introduce spoof detection techniques based on the time-to-live value found in the packet headers. From the spoofing analysis, we can deduce the distribution of operating systems that make up the sources of the attack.

The training of diffusion-based models for image generation is predominantly controlled by a select few Big Tech companies, raising concerns about privacy, copyright, and data authority due to the lack of transparency regarding training data. Hence, we propose a federated diffusion model scheme that enables the independent and collaborative training of diffusion models without exposing local data. Our approach adapts the Federated Averaging (FedAvg) algorithm to train a Denoising Diffusion Probabilistic Model (DDPM). Through a novel utilization of the underlying UNet backbone, we achieve a significant reduction of up to 74% in the number of parameters exchanged during training, compared to the naive FedAvg approach, whilst simultaneously maintaining image quality comparable to the centralized setting, as evaluated by the FID score.

Federated Continual Learning (FCL) is a emerging field with strong roots in Image classification. However, limited research has been done on its potential in Natural Language Processing and Tabular datasets. With recent developments in A.I. with language models and the widespread use of mobile devices, it becomes relevant to consider FCL’s capabilities in dynamic environments. Our
paper discusses and evaluates the applicability of FCL methods between the domains of Natural Language Processing Tabular Data with Image Processing as a baseline. We use Long-Short Term Memory (LSTM) models, DNN’s and LeNet-5 as models for Sentiment analysis, Tabular classification and Image classification. Through our experiments, we evaluate the average accuracy and backwards transfer of EWC, GEM, their federated variants and the state-of-the-art FCL method FedWEIT. With these methods, sentiment analysis and tabular classification tasks show that image classification reached over 17% higher average accuracy and achieved a 99.5% average increase in knowledge
transfer between tasks. Furthermore, we observe that non-federated continual learning methods on average reach higher accuracies than their federated
counterparts as well as the state-of-the-art methods

Federated learning enables training machine learning models on decentralized data sources without centrally aggregating sensitive information. Continual learning, on the other hand, focuses on learning and adapting to new tasks over time while avoiding the catastrophic forgetting of knowledge from previously encountered tasks. Federated Continual Learning (FCL) addresses this challenge within the framework of federated learning. This thesis investigates how FCL can be made vulnerable to Byzantine attacks (from unpredictable or malicious nodes), which aim to manipulate or corrupt the training process, compromising model performance. We adapt and evaluate four existing attacks from traditional federated learning in the FCL setting. Furthermore, we propose three tailored attacks for FCL are proposed based on the insights gained. Additionally, a novel attack called "Incremental Forgetting" is introduced, which specifically targets the incremental knowledge retention aspect of FCL. Our experimental evaluations of the attacks carried out against various FCL algorithms show that personalizing these towards FCL provides varying degrees of performance benefits, while the novel attack additionally exhibits evidence showing it may be more practical against real-world systems, strengthening its impact on the FCL community. This research contributes to the development of secure and resilient FCL systems to build better defenses against such attacks in the federated learning domain.

This thesis titled ´Study of 5G Roaming Security´ investigates the potential network vulnerabilities of 5G roaming reference points. The 5G Non-Standalone (NSA) is already being deployed in different countries across the world. With 3G becoming obsolete, mobile communication will primarily depend on 4G and 5G-NSA. In the later stages of the 5G rollout, 5G Standalone (SA) deployment will also take place gradually. Since 5G will support various use cases such as connected vehicles, smart farming, and smart healthcare, the number of connected devices will eventually increase. This would mean more data traffic generation compared to the LTE. In such a scenario, the privacy protection of the User Plane becomes highly significant. This thesis work provides a randomization-based security solution that would make the Standalone 5G reference points more robust to interception attacks.

The goal is to implement a negotiation-based randomization solution over N9, N3, and N32 as these reference points cross the HPLMN boundary. This solution will be a part of the GTP-U header and can be easily implemented by modifying the existing signaling procedures. Random bytes will be added to the GTP-U header before the start of the payload. The idea of adding randomization bytes has been extended to include TCP-based randomization and IMS-based randomization. The TCP-based randomization also includes two different algorithms for the addition of random bytes. After analyzing the User Plane security, the vulnerability analysis of SEPP was undertaken, to understand the vulnerabilities that can be a threat to the network infrastructure. A vulnerability assessment matrix was made and high-risk vulnerabilities were highlighted along with a few precautionary steps. The implementation details and architectural changes for the implementation of GTP and TCP-based randomization are provided. The randomization is useful in masking the signature distribution of an application's packet length and can be a powerful protection mechanism against data traffic analysis attacks.