Continuous Integration (CI) has become a fundamental practice in modern software development. Organizations increasingly adopt complex pipeline configurations to automate their build, test, and deployment processes. Well-optimized CI/CD pipelines offer significant benefits in deployment reliability, team productivity, and code quality. As these pipelines become more complex, defined by the number of jobs and steps in this paper, there is limited empirical evidence on how this evolution affects key performance metrics. This study addresses this gap by investigating the relationship between pipeline complexity and performance outcomes. By analyzing data from over 194 open-source GitHub repositories, we reveal that while increased complexity generally correlates with longer build durations, the impact on success rates is less direct. More importantly, we found that strategic modifications to pipeline configuration files (i.e., line changes) were frequently associated with significant performance improvements, including shorter build durations and fix times. This research provides guidance for practitioners. Rather than asking whether to increase or decrease pipeline complexity, our findings show that the focus should be on the method of change. We recommend prioritizing small, iterative maintenance activities, which consistently improve performance, over large-scale tool migrations, which have unpredictable outcomes. This approach enables teams to evolve their pipelines while mitigating the risks associated with growing complexity.

Property-based testing (PBT) allows developers to specify high-level properties that should hold for a range of inputs, which are then automatically generated by the testing framework. Despite its theoretical appeal, PBT is not widely used in the real world. To better understand how PBT is used in practice, we present a qualitative and quantitative study of 87 property-based tests written with the Hypothesis framework in seven widely used Python projects, including cpython, pandas, and jax.

Our analysis reveals that while PBTs are relatively rare in these repositories, they are often simple in structure and highly effective at expressing functional properties. The most common patterns include round-trip checks and comparisons against test oracles, which account for a significant portion of the test suite. We also observed a high rate of custom generator usage (39.1%), but no use of custom shrinking, suggesting strong defaults in Hypothesis. The dataset was variable in stylistic and structural choices, ranging from single-assertion tests to complex hardware-dependent cases.

This study provides new insights into the practical use of PBT in Python, expands on prior quantitative work with qualitative findings, and identifies concrete implications for improving testing frameworks and educational resources. We conclude with recommendations for supporting broader adoption of PBT and outline directions for future research, including cross-language comparison, automated annotation, and temporal analysis of test evolution.

Property-based testing (PBT) verifies software correctness by checking that specific properties hold across a wide variety of randomly generated inputs. Despite its apparent usefulness, we lack an overview of how PBT is utilized in the Java ecosystem. In this study, we investigated seven repositories using the jqwik framework. We analyzed 84 PBTs to understand the types of properties developers typically test, their use of generators and shrinkers, and the role of PBT in broader testing strategies. Our findings show that PBT remains a small part of most test suites, often representing less than 2% of the total number of tests. Developers tend to focus on mutation, invariant, and round trip as the properties they test, frequently using custom generators combined with filtering but never implementing custom shrinkers. We conclude that property-based testing is not being utilized to its full potential in Java projects and highlight areas for future research, including the impact of filtering, the potential of custom shrinkers, and the overall effectiveness of property-based testing.

Property-based testing (PBT) is a method of verifying software correctness in which a property, a statement about the behavior of the program which should always hold true, is verified with a large number of arbitrary inputs. While it is a powerful method, properties can be complex and fail on obscure inputs, making them difficult to reason about. We qualitatively analyze a large number of property-based tests (PBTs) written with Quickcheck for Rust from public repositories, in order to gather insights about how PBT is utilized in practice. We find that the majority of analyzed PBTs verify by comparing against a known correct implementation, composing inverse operations to form an identity, or asserting a desired contract about a system's state. These findings offer valuable direction to the development of PBT, aiming to tailor PBT frameworks to developer needs and make them more natural to work with.

Property-Based Testing (PBT) with QuickCheck has become a cornerstone of reliable software development in Haskell, yet there is little systematic understanding of how developers employ it in production quality libraries. In this study, we perform an empirical analysis of QuickCheck usage in nine representative open-source Haskell projects (aeson, attoparsec, containers, hashable, lens, megaparsec, pandoc-type, text and vector). We extract thousands of
QuickCheck properties and manually sample 217 of them to classify property types such as invariants, roundtrip checks, algorithm correctness, and idempotence. We also analyze all extracted properties to identify patterns in generator and shrinking strategies. Our Results show that invariants and test-oracle tests dominate the sampled properties, while custom generators are used in 55.8% of all properties and custom shrinkers in only 25%. We discuss how these patterns vary across different domains (data structures, text processing, optics, hashing) and highlight the tension between default and custom test infrastructure. Finally, we reflect on the implications for tool support, education, and future research particularly automatic generator inference and improved shrinking utilities to lower
the barrier to effective PBT in large codebases.

Continuous Integration (CI) practices have become central to open-source software (OSS) development, yet the relationship between branching strategies, merge habits, and CI performance remains underexplored. Understanding their role is crucial for explaining the variation in CI outcomes and for refining development practices. We empirically examine how branching models (feature-based vs.\ trunk-based) and merge characteristics (size and frequency) affect key performance indicators (KPIs).

Using a dataset of 565 GitHub repositories, we analyze both short-term trends and long-term evolution of development strategies. We find that while feature branching is strongly associated with higher delivery frequency and lower defect counts, trunk-based workflows (though rare) sometimes outperform in lead time and recovery. Similarly, frequent merges correlate with faster delivery and shorter lead times, regardless of size. A longitudinal subset reveals that projects shift toward feature-based development over time, but do not consistently adopt smaller or more frequent merges.

We also highlight methodological limitations in mining GitHub. Future research should incorporate longitudinal repository tracking and developer surveys to capture workflows that are invisible to snapshot-based analysis. This study contributes to a nuanced understanding of how code management practices shape CI outcomes in collaborative OSS projects.

Continuous Integration (CI) has become a standard practice for speeding up software development. However, the effect of comparatively slower artifacts, like documentation, on its performance is still unclear. Although documentation is often regarded as important, there is little data that connects documentation practices to key DevOps metrics. This study examines this relationship by looking at 670 open-source projects that use CI. We measured how documentation completeness, update frequency, and release notes affect delivery frequency, defect counts, and mean time to recovery. Our results show a "tipping point" where high documentation completeness greatly increases delivery frequency. We also found a "sweet spot" for update ratios between 20% and 55%, which relates to the lowest defect counts. On the other hand, we found no proof that long release notes improve recovery time. We conclude that the effectiveness of documentation depends more on quality and rhythm than on volume. This provides developers with practical, data-driven strategies to improve project performance.

The growing complexity of modern software systems, driven by larger codebases and evolving technologies, has amplified the need for effective collaboration in developer teams. Specifically, in open-source software (OSS) projects, where contributors often vary in background and engagement in the process, this complexity may introduce further collaborative challenges. As projects scale, coordination becomes increasingly difficult to maintain, highlighting the importance of understanding the socio-technical dynamics of effective development. While prior research emphasizes the role of collaboration, its influence on software delivery remains underexplored. In order to address the gap, this study examines how team characteristics, such as size and expertise, and communication practices, like interactions on issues and pull requests, relate to delivery efficiency in OSS projects. Based on an empirical analysis of 887 GitHub repositories, we found that team size and project expertise exhibit the strongest relationships with delivery size and frequency. Core contributor activity also shows positive but diminishing effects over time, while communication practices demonstrate no noticeable associations with release efficiency. These findings suggest that optimizing delivery in OSS projects may benefit from considering adaptive team structures and aligning CI/CD practices with the project stage and the evolving dynamics of the developer team.

Continuous Integration is a used extensively in modern software engineering for both proprietary and open-source projects. Many studies have studied its benefits and drawbacks, finding how it increases development productivity and stability. However, CI is a set of practices from static linting to calculating the code coverage of the underlying test suites. In order to choose whether to make use of that technology or to evaluate the overall performance of a project’s development, practitioners make use of certain measurements, DevOps metrics being one of the most significant ones. We aim to analyse the effects of testing strategies within the CI over a set of DevOps metrics. This is done by collecting over 5778608 executions of GitHub CI workflows that involve a test-running step from 476 open-source projects. We see that 69.48% of runs happen after a pull request or a push. In the end, we found that frequent CI test execution didn’t increase the project’s DevOps metrics, indicating that developers should try limiting the unnecessary execution of tests to save on resources. Further we see lack of Pearson statistical significance for the correlation between coverage in CI and the metrics in the smaller set of selected projects.

The proliferation of video recording devices and facial recognition technology has led to significant privacy concerns, as surveillance systems can capture and identify individuals without their consent. Traditional facial obfuscation systems, which introduce pixel-level perturbations to images, aim to protect privacy by preventing unauthorized facial recognition. However, these systems are vulnerable to inversion attacks, where attackers can reverse the perturbations to restore original images, compromising privacy. This thesis addresses these vulnerabilities by proposing HyperObf, a novel approach utilizing HyperNet technology to generate unique obfuscation networks for each user. HyperObf ensures that each user’s images are distinctly protected, making it challenging for attackers to reverse-engineer the obfuscations. Our experiments demonstrate that inversion attacks can significantly degrade the protection offered by static obfuscation systems, with restored images achieving face recognition accuracy close to that of original images.
In contrast, HyperObf effectively mitigates these attacks, reducing the attack success rate to 30% compared to 60% for existing methods. Additionally, HyperObf can generate 100 personalized MaskNets in 0.2 seconds using high-performance computing resources. These findings highlight the potential of HyperObf to enhance privacy protection against unauthorized facial recognition and inversion attacks in the digital age.

This research investigates the impact of multipath signals in UWB communications and explores their potential to improve localization accuracy of tags using the additional information captured in the Channel Impulse Response (CIR). While traditional localization typically relies on multiple anchors, this study focuses on using a single anchor and creating multiple Virtual Anchors (VAs) through MultiPath Components (MPCs). The research proposes algorithms such as likelihood, algebraic, and fminsearch methods, and tests their performance through simulations in environments with anchors, tags, and reflective surfaces.

After confirming the feasibility of using reflection multipath signals for localization through the simulations, experiments were conducted with NXP's SR150 UWB tags, which collect CIR data via two antennas. An indoor environment was set up with two MPCs and a strategically placed anchor. Data was collected from various tag locations and analyzed using the different localization methods.

The collected CIR data revealed additional reflections, notably from the ceiling and ground, which sometimes interfered with the desired MPC reflections. To address this, data cleaning techniques were implemented for the likelihood method, and the two antennas were used to approximate the AoA in the fminsearch and multilateration methods.

The final results demonstrated that the likelihood method achieved an average localization error of 26 cm, the fminsearch method 20.7 cm, and the algebraic method 24 cm. While the fminsearch method provided the best accuracy, it required data from two antennas, unlike the likelihood method, which only needed one. The fminsearch method showed improved accuracy near MPCs, though its performance declined as the tag moved further away. This suggests its potential for applications where AoA is large, as traditional methods become less accurate beyond 60-degree angles.

Indoor localisation is a well-researched topic and it is a challenge to improve the accuracy of existing techniques. In recent years, edge computing and federated learning have opened up new possibilities and challenges for indoor localisation. This thesis presents a federated implementation for spatial mapping of the network based on the RSSI signal between nodes. The proposed approach decentralises indoor localisation and produces multiple coordinate maps of the nodes in a network. The coordinate maps and edge data are from multiple node perspectives, which can improve the aggregation of the coordinates into a single map. The algorithm is tested in a BLE network and is compared with other current methods. In addition, a simulator is built that serves as a proof of concept for the actual implementation of the algorithm. The results of the simulator prove that the mathematical aspects are correct and that the algorithm consistently reproduces the same structure with a stable signal. Time-series analysis of RSSI measurements along with environmental factors unveils periodic noise components such as humidity and human presence, along with the notable influence of node placement and ambient noise. Consequently, the conventional RSSI-to-distance model proves inadequate in adapting to such noise sources. The spatial mapping algorithm helps in creating a 3D structure in a federated manner and provides a framework and location reference for future research that employs adaptable machine learning models. However, its localisation accuracy is still limited in comparison to other preexisting techniques.

Federated learning (FL) is a privacy preserving machine learning approach which allows a machine learning model to be trained in a distributed fashion without ever sharing user data. Due to the large amount of valuable text and voice data stored on end-user devices, this approach works particularly well for natural language processing (NLP) tasks. Due to many applications making use of the algorithm and increasing interest in academics, ensuring security is essential. Current backdoor attacks in NLP tasks are still unable to evade some defence mechanisms. Therefore, we propose a novel attack, the single-character strike to address this research gap. Consequently, the following research question is posed: What are the properties of the single-character strike in a language classification task? By experimental analysis the following properties are discovered: the single-character strike is undetectable against five state-of-the-art defences, has low impact on the global model accuracy, trains slower than similar attacks, relies on characters on the edge of the distribution to function, is robust within the global model, and performs best when close to convergence and with more adversarial clients. Emphasizing its imperceptibility and persistence, the attack maintains a 70\% backdoor accuracy after a thousand iterations without training and remains undetectable against: (Multi-)Krum, RFA, Norm Clipping and Weak Differential Privacy. By providing insight into the effective single-character strike, this paper adds to the growing body of work that questions whether federated learning can be secure against backdoor attacks.

Federated learning provides a lot of opportunities, especially with the built-in privacy considerations. There is however one attack that might compromise the utility of federated learning: backdoor attacks [14]. There are already some existing defenses, like flame [13] but they are computationally expensive [14]. This paper evaluates a version of differential privacy, where the Gaussian noise added to the aggravated model of the clipped updates is smaller than usually. This is often referred to as weak differential privacy or weakDP. This paper evaluates weakDP with different parameters to find if weakDP can be used as a defense for a language processing federated learning classifier against a backdoor attack.

Abstract— Federated Learning (FL) makes it possible for a network of clients to jointly train a machine learning model, while also keeping the training data private. There are several approaches when designing a FL network and while most existing research is focused on a single-server design, new and promising variations are arising that make use of multiple servers, witch have the benefit of speeding up the training process. Unfortunately single-server FL networks are prone to model poisoning attacks by malicious participants, that aim to reduce the accuracy of the trained model. This work showcases the inherent resilience of the multi-server design against existing state-of-the-art attacks tailored around single-server FL, as well as propose two novel attacks that exploits multi-server topology in order to reduce the required knowledge an adversary needs to obtain to carry out the attack, while still remaining effective. Main findings are as follows: In the event that the malicious party has compromised the entire network, existing single-server attacks are sufficient to completely prevent a model from training. If they are limited to knowledge available only within the local reach of their compromised clients, the effect is minimized to where the attacks might get mitigated without any defences being necessary. However in such cases a correlation can be observed between the location of the compromised clients and the effectiveness of an attack. The novel attacks proposed in this paper exploit this relation in order to remain sufficiently effective while requiring only the same amount of data necessary for the multi-server algorithm to function.

This thesis presents Screen Antenna - A Visible Light Communication (VLC) system that integrates data transmission and reception, with the conventional pixel display capability of RGB LEDs. The system is constructed with off-the-shelf components and runs on the Arduino Due microcontroller. The hardware and software have been designed with the objectives of scalability and flexibility. This research aims to optimize the system's performance in terms of throughput, error rate, and display quality.

To evaluate the system's performance, extensive experiments were conducted under different lighting conditions, PWM frequency and Rx-Tx distances. The experiments focused on achieving the maximum achievable throughput while maintaining a low Bit Error Rate (BER). Through careful optimization of the system parameters, a maximum throughput of 4.4kbps was achieved. Furthermore, the system consistently maintained a Bit Error Rate of less than 0.005\%, ensuring reliable and error-free data transmission.
In addition to data transmission, the impact of data transmission on the display's visual quality was investigated. The Delta E value, which quantifies the observed colour difference between two pixel values, was used as a metric for display quality. The data rates were calibrated such that the display operated with a Delta-E value of less than 1 in more than 90\% of the colour spectrum, indicating minimal colour distortion during data transmission.

This study contributes to the field of VLC by showcasing a practical implementation that combines scalability, visual quality, and performance. Future research directions could focus on expanding the display size and enhancing the data rate by means of more sophisticated hardware.

End-to-end Automatic Speech Recognition (ASR) systems improved drastically in recent years and they work extremely well on many large datasets. However, research shows that these models failed to capture the variability in speech production and have biases against the variant caused by the regional accented speech. Moreover, ASR research on regional accents is primarily done in languages used by a large population, like English and Arabic, and the effect of regional accented speech on E2E ASR systems in non-popular languages is still unknown. It is important to know the effect of regional accented speech on E2E ASR systems as it helps researchers to build an inclusive E2E ASR system. In this project, I aim to mitigate the biases against regional accented speech. I select standard speech and regional accented speech from CommonVoice's French and German datasets. I combine the state-of-the-art Conformer Recurrent Neural Network Transducer model with Multi-Domain Adversarial Training (MDAT) to boost the performance of regional accented speech while not hurting the performance of the standard speech. Moreover, since the regional accented speech is typically low-resourced, I study the amount of data required for effective MDAT, as well as the effect of different domain classifiers on the performance of Multi-Domain Adversarial Training. Experimental results show that MDAT can mitigate the biases against regional accented speech in both French and German. The best model in French reduces the bias by around 12% and the best model in German reduces the bias by around 7%. Additionally, MDAT is an effective method for bias mitigation as it can achieve similar performance as the MDAT model trained with the full dataset using only a small amount (e.g. 30 minutes) of untranscribed regional accented speech. Finally, different domain classifier architectures were found to have similar effects on the results of MDAT, thus there is no significant differences among the domain classifier in this project.

Due to recent developments in DNA sequencing technology, there is a growing abundance of available genomic data. To process this information for use in fields such as healthcare and forensics, raw sequencing data have to be processed using computationally intensive algorithms. Currently, one of the major bottlenecks in this processing pipeline is the alignment step, which makes use of dynamic-programming algorithms. To reduce computation times, numerous solutions have been proposed aimed at reducing the execution time of the alignment step. This is done either by accelerating alignment itself using hardware accelerators and heuristics or by reducing the amount of input data through the use of pre-alignment filters. The algorithms associated with the latter solution are less computationally intensive than DP-based alignment, which reduces the end-to-end alignment time.

Currently, pre-alignment filters are effective to the point where the alignment bottleneck is shifted to the filtering step. Therefore, the filters are accelerated on hardware solutions such as GPUs and FPGAs. While these solutions show orders of magnitude improvement in execution times, they are insufficient for removing the filtering bottleneck entirely. The performance of these hardware accelerators is limited by the rate at which data can be supplied. As a solution, we propose a CIM-based accelerator to reduce data-movement overheads between the host device and the accelerator. Additionally, this architecture makes use of emerging non-volatile memories to perform Boolean operations directly within its memory elements. In doing so, it can exploit parallelism in the algorithms to achieve higher throughput.

In this work, we explore commonly found operations in existing pre-alignment filters and devise ways to implement them on the CIM-architecture. The proposed architecture is flexible in supporting multiple pre-alignment filters and a wide range of input data. The functionality of the architecture is verified through simulation and its effectiveness is tested using real data sets.

Using this architecture, we can achieve improvement in end-to-end execution time over the state of the art ranging from 7.2x to 119.6x for the evaluated data sets, while also achieving a reduction of up to 59% and 79.7% in chip-area and power consumption, respectively.

Furthermore, the provided work offers a platform for the development of future pre-alignment filtering algorithms to further improve performance.

The considerable increase in the number of devices needing connectivity, such as mobile phones and Internet of Things (IoT) devices, has led to an exponential rise in data volumes during the last years, that will surely continue over the next decade. Therefore, it will be increasingly challenging to provide sufficient RF resources. A novel alternative to RF communications is Visible Light Communication (VLC). VLC is a communication technology that uses visible light as an information carrier. The use of VLC for indoor applications has been rapidly growing during the last years – Light Fidelity (LiFi) technology is an example of VLC application – with photodiodes being the most widely used receiving devices. However, looking at both indoor and outdoor communication, photovoltaic (PV) cells represent a relevant alternative for detecting the information. One of the advantages of using a PV cell as receiver is the huge sensitive area for detection of the information that simplifies alignment between transmitter and receiver.

Different light sources can be used in VLC. Typically, either LEDs or LASERs are considered, depending on the characteristics of the link (such as distance, type of receiver, indoor/outdoor application). These light sources differ in terms of spectrum, directionality, optical power density and bandwidth. The performance of the whole VLC link strongly depends on the characteristics of the light source, since it affects the ability of the receiver, such as a PV-device, to detect the information when it overlaps with the ambient light, that can reach very high values, especially in outdoor applications where the ambient light is the sunlight. Therefore, the modelling and analysis of the performance of different light sources in a PV-based VLC link will pave the way towards the realisation of a PV-based communication system of the future; and it is the focus of this thesis project.

The project goals were achieved by first reviewing the characteristics of light sources, to understand their advantages and drawbacks in (PV-based) VLC. This was followed by the development of models of the VLC data-link, with a focus on the light source, which took into account various factors such as the type of light source, its location relative to the receiver, and its dynamic behaviour. This was followed by the realisation of a test setup, to characterise different light sources, and the models were then used to simulate the light distribution from the actual light sources. Finally, the framework was used to simulate a LED-based solar simulator and an outdoor VLC data-link.