Recent proposals to deploy terrestrial 5G/6G networks in the 12 GHz downlink band (12.2–12.7 GHz) have raised concerns about interference to incumbent non-geostationary orbit (NGSO) fixed-satellite service (FSS) systems such as SpaceX’s Starlink. This thesis investigates the technical feasibility of spectrum coexistence between the downlink of an NGSO FSS network and a prospective terrestrial 6G mobile network in this band. Two influential but conflicting prior studies, one by RKF Engineering on behalf of terrestrial stakeholders and one by SpaceX, are first replicated and analysed to identify the modelling assumptions that drive their divergent conclusions. Building on this analysis, a revised modelling approach is developed, accelerated using general-purpose GPU computing and incorporating more realistic deployment scenarios, updated propagation and clutter models, and refined NGSO FSS and mobile-network parameters.

Simulation results indicate that simultaneous operation of NGSO FSS downlinks and 6G mobile networks in the 12 GHz band does not satisfy the ITU-R NGSO FSS protection criteria in urban environments. In both macrocell and small-cell terrestrial deployment scenarios, a large fraction of user terminals in urban areas experience interference levels exceeding the applicable INR protection criterion, which is consistent with SpaceX’s assessment and inconsistent with the optimistic predictions of the RKF study. Small-cell architectures reduce exceedance in suburban and rural regions but leave urban exceedance largely unchanged. Co-channel interference remains severe wherever dense terrestrial deployments coincide with NGSO FSS user terminals. These findings suggest that sharing the 12 GHz downlink spectrum between NGSO FSS and a terrestrial 6G network would entail degradation risk for satellite broadband in populated areas, and motivate further research into alternative band arrangements, coexistence concepts (including uplink-focused use), and spectrum-allocation strategies.

Text-to-image diffusion models have advanced significantly in recent years. Different models show strong performance across various generation tasks. Choosing the right model is becoming increasingly important since no single model consistently outperforms others in all cases. However, existing model selection approaches are typically evaluated only at the dataset level. Such evaluation overlooks prompt-level variation, where different models may excel on different prompts. In this thesis, we investigate diffusion model selection during inference. The goal is to pick the best model for each individual prompt. We first examine the online setting, where model selection occurs adaptively during deployment. In this context, we create a framework for online diffusion model selection and test it against recent methods from the literature. Our findings show that this approach outperforms existing online selection strategies, highlighting the benefits of prompt-aware model selection. In addition to the online setting, we present an offline approach to diffusion model selection, where decisions are made without online interaction. Overall, this thesis claims that diffusion model selection should be viewed as a prompt-level decision rather than a dataset-level comparison. By exploring both online and offline settings and providing empirical results alongside detailed ablations, we aim to promote a more practical and adaptable approach to diffusion model selection.

Long distance Haptic Bilateral Teleoperation (HBT) is used in applications such as surgery, training, remote operation, and disaster relief. One of the main challenges in these systems is communication delay. When force feedback relies directly on sensors in the remote environment, increasing delay quickly makes the system unstable and hard to control. To overcome this, previous approaches generated force feedback using a local simulation of the remote environment, while providing visual feedback through live video [20]. This reduces timing constraints, but only works when the simulation accurately represents reality. When the robot encounters objects that are not modeled and not visible to the operator, unpredictable forces occur and the system no longer behaves correctly.
This thesis investigates how such unpredictable forces can be handled in delayed HBT. We introduce a method that keeps predictive force feedback through simulation, while correcting the simulation using sensor data from the robot. A small 3D printed attachment was developed to detect unexpected contact events. These measurements are sent back to the operator side and used to update the virtual environment, allowing force feedback to be generated even for unknown objects. The approach was evaluated in a user study with 13 participants under varying delays and visibility conditions. The results show that when visibility is limited, reaction based feedback can be used instead of prediction based feedback. The findings indicate that combining simulation based prediction with remote sensing offers a practical solution for dealing with unpredictable forces in long distance HBT.

Event-driven neural network accelerators achieve superior energy efficiency by processing only meaningful data events, yet existing design space exploration tools lack support for their asynchronous execution characteristics. This thesis introduces AeDAM (Event-Driven Architecture Mapping), a specialized framework for systematic exploration of event-driven accelerator architectures.

AeDAM transforms traditional synchronous mapping methodologies into event-driven configurations through intelligent Loop Order Memory Access scheduling and specialized analytical cost models for asynchronous dataflows, targeting energy-delay product optimization.
Experimental validation using the SENECA neuromorphic architecture demonstrates substantial improvements: 2.5× faster exploration times, 13-52% latency reductions across VGGNet layers, and 12× energy-delay product improvements. Optimal configurations feature 512KB SRAM capacity and multi-dimensional processing element arrays.

AeDAM establishes a foundation for systematic exploration of energy-efficient event driven computing systems targeting edge applications.

This research explores the feasibility of implementing lane detection on lightweight microcontrollers using a combination of traditional image processing and compact machine learning methods. With the aim of enabling real-time inference under strict hardware constraints, several models were trained and evaluated against a custom image processing pipeline. Each approach was tested for accuracy, speed, and resource usage on the Raspberry Pi Pico 0 microcontroller. While these solutions fall short of cutting-edge accuracy and cannot process as much information as state of the art models, their low cost, minimal power consumption, and real-time performance highlight their potential. These findings suggest that lightweight lane detection is a viable direction for further research in embedded autonomous systems.

Real-time traffic sign recognition on microcontrollers introduces challenges due to limited memory and processing capacity. This study investigates the trade-offs between model size, classification accuracy, and inference latency within hardware constraints. We present an efficient network architecture called AykoNet with two variants: AykoNet-Lite, prioritizing model size and inference latency, and AykoNet-Pro, prioritizing classification accuracy. We trained AykoNet on the German Traffic Sign Recognition Benchmark (GTSRB) and specifically optimized it for deployment on the Raspberry Pi Pico microcontroller. AykoNet-Lite delivers 94.60% accuracy with only a 36.80KB model size and 55.34ms inference time, while AykoNet-Pro achieves 95.90% accuracy with an 80.18KB model size and 87.13ms inference time. Our approach demonstrates the effectiveness of domain-specific preprocessing and architectural design, class-aware data augmentation, and the strategic use of depthwise separable convolutions. These results validate the feasibility of real-time traffic sign recognition in resource-constrained embedded systems. Specifically, AykoNet-Lite strikes an optimal balance between model size, classification accuracy, and inference latency for practical deployment in autonomous navigation applications.

Visible light positioning (VLP) enables accurate indoor localization by leveraging a dense deployment of LEDs in future lighting infrastructure, but its widespread adoption is hindered by two key challenges: the need for densely sampled fingerprint datasets and performance degradation due to LED aging or failure. In this work, we propose a VLP framework that reduces reliance on dense fingerprinting and remains robust over time without requiring manual re-fingerprinting. Using a dataset acquired from the DenseVLC testbed, we evaluate preprocessing techniques that enhance positioning accuracy under noisy received signal strength (RSS) measurements. To address long-term reliability, we introduce a simulation framework that models LED degradation and sudden failures. Most importantly, we present an online learning approach that dynamically adapts the positioning model in response to environmental and infrastructure changes.
In our simulations, this approach maintains the original level of accuracy despite aging effects. In some cases, it yields up to a 95% improvement when evaluated over longer timespans. Furthermore, our preprocessing contributions have led to a 30% improvement to baseline performance without aging. Our results demonstrate a path toward scalable, self-sustaining VLP systems suitable for real-world deployment.

Visible light positioning (VLP) systems are a promising solution for indoor positioning, utilizing light-emitting diodes (LEDs) as transmitters and photodiodes (PDs) as receivers.
A received signal strength (RSS) based VLP system's accuracy is heavily dependent on the density of collected fingerprints, being a very labor-intensive process.
In this study, we focus on RSS fingerprints to achieve centimetre level positioning accuracy, while addressing the challenges of labor-intensive fingerprint collection and deployment on resource-constrained devices like the Raspberry Pi Pico microcontroller.
We found different neural network architectures using Neural Architecture Search (NAS) to optimize the VLP system, which achieve on average $12mm$ positioning error with low inference latency around $50ms$ on the Raspberry Pi Pico.

This work investigates the feasibility of performing monocular depth estimation on highly resource-constrained hardware, specifically the Raspberry Pi Pico Zero microcontroller. In contrast to existing approaches that rely on large convolutional networks and high performance devices, this study explores a set of custom lightweight encoder-decoder architectures, including one inspired by L-ENet, L-EfficientUNet, μPyD-Net, and an LSTM-μPyD-Net combination, designed to operate within strict memory limits. These models were trained on a preprocessed KITTI dataset, with either LiDAR depth maps or SGM (Semi-Global Matching) dense depth maps, and evaluated in terms of accuracy, model size, and real-time inference performance. Results demonstrate that meaningful depth prediction is achievable on microcontrollers, paving the way for low-cost autonomous navigation systems and broader applications of TinyML in embedded robotics, with SGM proving to be the best preprocessing technique, and the LSTM-μPyD-Net having the best accuracy when trained on the full Train split of the KITTI dataset.

Liquid Crystal Reconfigurable Intelligent Surfaces (LC-RIS) are crucial for future millimeter-wave and terahertz wireless systems due to their ability to dynamically control electromagnetic waves. As LC-RIS requires very fast operation, an appropriate voltage supply with good response time must be developed to adjust each element independently. In this work, we present the design and implementation of a custom voltage supply tailored for LC-RIS applications. Our system integrates four daisy-chained LTC2688 digital-to-analog converters (DACs), providing 64 independently adjustable voltage outputs. We achieve a worst-case update time of 273 µs for all outputs, using a Raspberry Pi Pico as the SPI master with minimal C code. The voltage supply meets critical system demands, including the generation of 1 kHz square waveforms, a voltage range of ±10 V, and sub-millisecond response times, while maintaining ease of control through a programmable interface. We also developed a dedicated on-board power supply to meet the LTC2688’s multiple power rail requirements, enabling the entire system to operate from a single external input between 3.7V and 18 V. Furthermore, our design is scalable, allowing multiple boards to be daisy-chained for larger LC-RIS arrays. Our solution demonstrates a compact, fast, and scalable voltage control system suitable for high-performance LC-RIS platforms.

Liquid-crystal reconfigurable intelligent surfaces (LC-RIS) need hundreds of stable bias voltages, yet most existing controllers are slow, expensive, or hard to scale. This thesis prototypes a lean 64-channel driver that combines one 1 kHz NE555 square-wave source with per-channel amplitude control via daisy-chained AD5263 digital potentiometers and OPA4197 buffers. Each channel supplies a continuous ±10 V swing in 256 steps, and a Raspberry Pi Pico streams updates in 31 μs, fast enough not to limit the LC’s millisecond-scale response. The work is still a proof of concept: it has been validated only with 64 channels on the bench, long-term drift and true large-panel scaling remain open questions. Nevertheless, the prototype points toward a practical, low-cost path for future LC-RIS control.

WiFi sensing has shown great promise in applications such as activity recognition, human identification, and health monitoring. However, models trained on Channel State Information (CSI) data often suffer from poor generalizability due to high variance and inconsistent reporting practices, especially in small-data regimes. Despite increasing attention to deep learning-based approaches, the community lacks standardized guidelines on handling variance and overfitting across architectures and datasets. In this work, we first review existing robustness strategies employed in related fields to identify techniques suitable for WiFi sensing. Based on this survey, we systematically benchmark five representative methods—stability training, mixup, coupled weight decay, early stopping, and label smoothing—selected for their theoretical grounding and prior success in mitigating variance. Our evaluation spans three model types (LeNet, LSTM, CNN+GRU) and two CSI datasets (Widar, NTU-Fi), using stratified 5-fold cross-validation with repeated trials to ensure reliable variance estimation. Results show that (1) stability training consistently improves moderately performing models, (2) coupled weight decay is especially effective for LSTMs, and (3) combining techniques can harm performance in near-saturated scenarios. Our findings offer actionable, architectureand dataset-specific guidelines for improving robustness and reproducibility in WiFi sensing research.

Wi-Fi sensing poses a serious threat to privacy due to its passive and covert nature. Nonetheless, the field of defenses is largely underdeveloped. This survey draws from the vast field of radar systems and their state-of-the-art countermeasures to discover potentially new Wi-Fi sensing defenses. The study identifies four particularly promising approaches: false target generation for deceptive jamming, reconfigurable intelligent surfaces (RIS) for dynamic signal manipulation using metasurfaces, encrypted waveform design for secure transmission, and hybrid region-based techniques for spatial access control. By transferring insights from radar systems to the context of Wi-Fi sensing, this work lays the groundwork for the future development of practical and resilient privacy-preserving defenses.

Knowing the floorplan of an incident site beforehand allows first responders to operate quicker and more efficient. This thesis explores the potential of using robotic swarms for environment mapping, with the goal of deploying these swarms to create maps before emergency personnel arrive on the scene. We present BICLARE, a lightweight collaborative algorithm for robust exploration. Inspired by ant colony behaviour, specifically how they use pheromones for communication and navigation, BICLARE implements a confidence model to determine the occupancy of cells within a map. Target selection considering travel time and estimated battery power ensures its efficiency. Including computation-saving parameters ensures it lightweight execution, enabling it to work on inexpensive hardware. The performance of the algorithm was evaluated through a series of simulated experiments in a variety of environments, proving it can generate accurate maps with adequate coverage in noisy, volatile environments. A real-life experiment demonstrated that it can successfully run on low-cost hardware in a real-world experiment.

We present a passport-level trust token for Europe. In an era of escalating cyber threats fueled by global competition in economic, military, and technological domains, traditional security models are proving inadequate. The rise of advanced attacks exploiting zero-day vulnerabilities, supply chain infiltration, and system interdependencies underscores the need for a paradigm shift in cybersecurity. Zero Trust Architecture (ZTA) emerges as a transformative framework that replaces implicit trust with continuous verification of identity and granular access control. This thesis introduces TrustZero, a scalable layer of zero trust security built around a universal ”trust token”- a non revocable self-sovereign identity with cryptographic signatures to enable robust, mathematically grounded trust attestations. By integrating ZTA principles with cryptography, TrustZero establishes a secure web-of-trust framework adaptable to legacy systems and inter-organizational communication.

Modern ecological monitoring needs long-term, non-invasive observation of wildlife in remote areas, but traditional methods either store raw audio for offline analysis or depend on power-hungry, networked infrastructure. This creates three problems: high energy cost, high data volume and no real-time awareness in the field. This thesis presents ChirpAlert, a fully autonomous low-power acoustic sensor platform that records and classifies environmental sounds directly on embedded hardware. The system is built around a custom low-cost embedded platform node with an analog, ultrasonic capable microphone front-end, on-device machine learning inference, environmental sensing, GPS and long-range LoRa connectivity for transmitting live alerts. ChirpAlert runs an optimized audio event detection pipeline that converts raw audio into Mel-based time-frequency features and applies a quantized neural network for bird vocalization detection on the sensor itself. A training and deployment framework is introduced that automatically assembles a task-specific dataset for a target species and region, performs hyperparameter and preprocessing optimization and converts the trained model to be capable of running on embedded hardware. The final platform operates at approximately 56 mW active power while running on-device inference, enabling long-term deployment. Field tests confirm that ChirpAlert can recognize target bird sounds with 88.8% accuracy during deployment while correctly labeling 98.1% of the other sounds as not the target species, demonstrating the effectiveness of the proposed ChirpAlert hardware and audio event detection pipeline.

With the rise of the artificial intelligence starting in the late 2010s there has been a great increase of demand for neural networks. It seems AI is omnipresent these days, whether it is OpenAI releasing the famous or infamous large language model chatGPT, Google following by adopting a similar approach with integrating Artificial Intelligence (AI) into it's search engine or help desks becoming entirely controlled by AI. It is clear, AI is here to stay. NVIDIA normally a company thought of selling Graphical Processing Units (GPUs) for gaming purposes has shifted its focus to making GPUs used for running neural networks, this shift has shown its returns as it is now the most valuable company in the world.

The increased interest in AI does leave a lot of academia wondering: What actually is artificial intelligence and especially what are neural networks. With the introduction of deep learning, a machine learning method that searches for features in a dataset, neural networks have become more and more abstract. Developing an appropriate solution to better interpret, visualize and adjust the inner workings of a neural network is not just needed to keep track of the underlying dynamics of a neural network, but also to allow the user to go outside of the standardized framework to fine tune a network.

This thesis explores the functionality of neural networks, more specifically spiking neural networks, and describes the workings and functions of the created simulator. Multiple helper functions, that allow the user to both visualize and adjust the network outside of the standardized framework, are introduced that distinguishes the created simulator from the already present SNN simulators. Afterwards, five small simulations are done to validate the basic functionality of the simulator. Following these simulations the simulator is validated using two large datasets and comparing the simulator with the results from a well known spiking neural network library SLAYER. The first validation is done with a smaller data set containing hand written images of the numbers zero to nine. The second validation is done with a larger data set containing multiple different kind of performed hand gestures. Both validations show that the simulator has a high accuracy with the maximum mean square error between the SLAYER output and the simulators output being less than 2%.

This thesis researches the security of firmware images in the Internet of Things (IoT) and embedded devices. We present an open-source tool, Embedded Binary Analysis Tool (EBAT), designed to analyze cross-architectural firmware image security context. EBAT consists of various modules capable of discovering outdated software for various libraries, particularly on cryptographic libraries, and detecting Common Vulnerabilities and Exposures (CVEs), focusing on firmware's cryptographic libraries. It also detects exploit mitigation techniques on firmware's image binaries and discovers credentials and passwords with a focus on private keys embedded in the firmware image. Additionally, EBAT identifies Application Programming Interfaces (APIs) cryptographic misuses through static taint analysis (backward tracking) on cross-architectural binaries. We presented a total of 18 well-defined cryptographic rules and a list of 733 function calls with more than 1,600 function arguments, applicable in static taint analysis to check the possibility of cryptographic misuses based on 10 well-used open-source cryptographic libraries APIs. EBAT's static taint analysis provides a powerful framework for detecting the possibility of cryptographic misuses in cross-architectural binaries, making it a valuable tool for identifying and addressing vulnerabilities in cryptographic implementation in firmware images.

Using EBAT, we conducted a large-scale analysis of over 36,000 firmware images publicly crawled from the Internet and successfully unpacked over 60% of them. The created dataset of firmware images includes more than 5,000 different products across 33 vendors, spanning more than 20 years and a plethora of various device types. Our findings show that ARM and MIPS are the most prevailed CPU architecture in the IoT/embedded industry. We compared identical binaries across all vendors, revealing a significant percentage of similar binaries used across different vendors' firmware images. Our analysis of firmware binaries reveals a notable absence of exploit mitigation techniques in IoT/embedded firmware images, and we present many firmware images containing private keys, posing potential security threats. Additionally, versions of open-source cryptographic libraries used in firmware images are identified, and the CVEs of the cryptographic libraries are evaluated. Two real-world case studies on hard-coded credentials demonstrate the significance of the large-scale attack presented in this thesis. Hashed passwords, predominantly using outdated algorithms, have also been discovered, and several have been cracked.

The main goal of EBAT is to identify cryptographic misuses in cross-architectural binaries. By applying static taint analysis (backward tracking) to well-defined APIs on specific functions and arguments for 10 open-source cryptographic libraries, we can identify potential violations of cryptographic rules. This analysis was executed on over 1.4 million binaries, revealing that approximately 50% of examined firmware images violated at least one cryptographic rule. Various case studies on real-world vulnerabilities in firmware images are presented, including recent CVEs that are found in various vendors' products. Executing EBAT on those vulnerable firmware images, we tested the effectiveness of our tool to evaluate the automatic capturing of these known vulnerabilities. In addition, performing large-scale analysis on an extensive corpus of firmware images allows us to discover that other firmware images are affected by these known vulnerabilities, in some cases also across various product lines not covered on the public CVEs reports.

In conclusion, EBAT is a valuable resource for researchers working on firmware security. Its automated analysis process, comprehensive modules, and ability to discover possible vulnerabilities, cryptographic misuses at a binary level, and other security weaknesses make it a powerful tool for identifying and mitigating security risks in IoT/embedded devices.