Detecting money laundering in financial transaction data is a task where graph neural networks (GNNs) have shown strong potential. Such data is naturally represented as a directed multigraph, since two accounts, each represented as a node, may exchange many separate payments, each forming a distinct edge with its own amount, currency, and timestamp. Preserving these parallel edges, rather than collapsing them into a single connection, retains the fine-grained structure that allows for distinguishing laundering behaviour from ordinary activity. Yet these models also introduce a new vulnerability, as an adversary could manipulate the transaction graph to alter the neighbourhood of a suspicious account such that the GNN misclassifies it as benign. Existing adversarial robustness research operates on the adjacency matrix, which records at most one edge per node pair and therefore cannot represent the parallel transactions between two accounts that this task depends on. Multigraph GNNs therefore lack both a framework for evaluating robustness under structural perturbations and defences against such perturbations.

This thesis extends adversarial robustness analysis to multigraph GNNs through three contributions. First, it reformulates GNN message passing and attack optimisation over the incidence matrix instead of the adjacency matrix, yielding the first gradient-based structural attack that retains multi-edge structure. Second, it introduces unnoticeability loss terms that constrain perturbations to maintain the graph's statistical fingerprint, including the frequency of characteristic patterns such as short transaction cycles, keeping the attack statistically plausible and unnoticeable at the macro level. Third, it scales the framework to large networks with projected randomised block coordinate descent. On the IBM synthetic anti-money laundering dataset, learned attacks substantially reduce detection accuracy compared to non-learnable perturbations, and adversarial training recovers robustness, showing that multigraph GNNs are both vulnerable to structural manipulation and defensible against it. ... Read more

Automated dependency-update tools such as Dependabot help projects keep dependencies secure and compatible, and CI results give maintainers practical evidence for deciding whether such updates can be merged, delayed, or require closer inspection. This evidence is useful only if the same update state tends to produce the same observable outcome when evaluated again. Prior work has studied reproducible dependency-update failures in ecosystems such as Java/Maven; this thesis examines whether similar CI-like evidence can be obtained for JavaScript/npm projects, where lockfiles, package-manager behavior, browser tooling, native addons, and peer dependencies may introduce instability.

Building on prior work on automated dependency updates and reproducible
dependency-update benchmarks, this study evaluates 2,777 npm dependency-update pull requests from 3,083 Dependabot candidates. The results show that most outcomes are reproducible, but 152 experiments (4.8\%) and 111 PRs (4.0\%) arenon-reproducible. Instability is concentrated in tests and appears more often in browser-coupled, native-addon-risk, peer-dependency-risk, and Yarn-based projects. These findings suggest that CI remains useful evidence for npm dependency-update decisions, but practitioners and researchers should treat single-run outcomes as incomplete evidence in instability-prone settings. ... Read more

Modern software depends heavily on third-party packages, and the dependency-update pull requests (PRs) that maintain them are mostly opened by bots. Prior empirical studies measure how individual projects respond to these PRs. They usually pre-filter to Dependabot at discovery and read each PR's outcome from its GitHub state alone. This leaves a question unanswered: when the same upstream version pair reaches many projects, do they decide alike?

Answering it requires a pipeline that keeps every PR source and that can tell a real decision apart from bot workflow. We built a pipeline covering five ecosystems (Maven, npm, Cargo, pip, Go) and four PR sources (Dependabot, Renovate, other bots, humans). It classifies each parsed dependency change along five dimensions, validated at macro-F1 >= 0.85 against a 500-PR / 1,000-change gold set.

With the classified changes we measure cross-project alignment with a normalised-entropy score A(t) over (package, v1, v2) triples seen by at least five deciding projects, of which we found 1,898 candidates. Our first result is methodological. Around three-quarters of the PRs that look closed-unmerged are bots superseding or regrouping their own PRs. This can be determined from the comments under the PR. These account for 26.3% of all parsed PRs, far more than the 9.6% that are genuinely closed-unmerged.

Once we separate them, overall alignment is high (mean A(t) = 0.70, median 0.86). The sharpest disagreement is on security-flagged updates (A(t) = 0.44 against 0.74 for non-security): projects agree on them the least, not the most. Two consequences follow. The merge/close ratios reported by prior bot-PR studies mix project decisions with bot workflow at roughly one to three in our data, so they cannot be read as accept/reject rates. And crowd-based compatibility heuristics are weakest exactly where the stakes are highest: security-flagged updates. ... Read more

Dependency management is a critical, difficult-to-automate task in software engineering. Researching automated dependency management requires reliable, reproducible datasets of dependency updates across ecosystems, but existing datasets fall short: they cover only specific update types (e.g. breaking updates) and log outcomes rather than the causal factors behind pull request (PR) acceptance and build outcomes. Closing this gap would let researchers determine not just whether a dependency update PR succeeded, but why — information needed to build dependency management tools that developers can trust. As a first step toward this, we construct a categorisation model that describes the code changes within accepted dependency update PRs on a commit level, developed using an established taxonomy-development methodology, and build a regex-based tool that automates this categorisation with 86\% accuracy on hand-labelled data. Our results show that simple, deterministic techniques can reliably support transparent, automated change categorisation — a building block for future causal datasets of dependency updates. ... Read more

This study introduces an end-to-end pipeline for automatically mining, reproducing and classifying the results of dependency update pull requests in the Rust/Cargo ecosystem, Colloquially named DURP (Dependency Update Reproduction Pipeline), its main purpose is to assist researchers in two closely related themes in the field of software supply chains, namely breaking dependency updates and build failures. We adopt techniques from both aisles into a novel combined approach that attempts to study both phenomena at once. The pipeline is implemented for Rust/Cargo, a relatively new and promising programming language ecosystem that has rapidly grown since its creation [27]–[29]. As empirical proof we’ve applied DURP to produce a dataset of 20,536 dependency pull request repro- duction attempts, comprised of 2 cohorts: one mined through DURP’s tooling, a monthly-stratified sample of 16,111 candidates spanning 2020–2026, and an n = 4,425 sample extracted from a previous empirical study on automated pull requests [8]. Our work introduces a novel build environment selection policy under which both cohorts reproduce at statistically indistinguishable rates (59.5 % each; 95 % Wilson CIs 58.8–60.3 and 58.1–61.0 for live mine and rebatchi respectively). Previous work on breaking dependency updates relies on pre-filtering for breaking dependency candidates. In contrast, by being unselective about pipeline status during sampling, DURP discovers breaking up- dates naturally—978 breaking in total, 8.0 % of the 12,155 deter- minable reproductions—with the live mine breaking significantly more often than rebatchi (8.8 % versus 5.3 % of each cohort’s reproductions; non-overlapping 95 % CIs 8.3–9.4 and 4.5–6.2). Further, the monthly stratification across the 2020–2026 body, with a sparser 2018–2019 tail (2018–2026) allows us to study how both breaking updates and build reproducibility vary through time, revealing several notable findings about the maturity and nature of the ecosystem. Finally, we defend a 6-class taxonomy for build reproduction failures compiled from existing literature and extended for the ecosystem, validated by three independent coders at an inter-rater agreement of Fleiss κ = 0.613 ... Read more

This thesis investigates the observable structure of dark web DDoS-for-hire ecosystems to support operational disruption planning and presents an automated CTI pipeline for repeatable measurement. The pipeline integrates seeded discovery, Tor-based crawling, harmful-content filtering, deterministic indicator extraction, and LLM-assisted structuring. In the fixed reporting snapshot, 389,006 raw pages were reduced to 304,949 sentry-passed pages, 3,919 extracted pages, and 301 DDoS-classified URLs across 59 domains. Linkage analysis in this dataset found fragmented but actionable cross-domain signals, while market profiling showed sparse pricing and attack-vector fields. These results support triage and monitoring, rather than full attribution or complete market quantification. The main contribution is a transparent and reproducible measurement workflow that supports independent replication and keeps evidence interpretation within explicit confidence boundaries. ... Read more

In the domain of intraoral hydration sensing to monitor oral health, there is sparse research on continuous capacitive sensing. Most devices can only perform a single point of measurement, making it unsuitable for wearing long-term. This paper aims to research the possibility of expanding the Densor developed by Dsouza et al. [4], with a capacitive sensor to measure hydration in the mouth, by testing it on NaCl-demineralised water solutions. The obtained data does not yet show any strong correlation between salt values and measured values, suggesting the sensor might not be suitable for this purpose, but the data is yet insufficient to draw any definite conclusions. ... Read more

Continuous and prolonged breathing rate monitoring is a challenge. Existing technologies have limitations such as bad user comfort over time or interference from the experimental environment. Intraoral sensors offer a novel approach, measuring at a position directly connected to the lungs and breathing cycles. The objective is to develop and evaluate a breathing rate estimation pipeline using pressure and acceleration sensor measurements.The Densor (dental sensor) is an intraoral sensing device that generated the required data.Participants performed controlled experiments with varying breathing rates and body positions.Pressure and acceleration measurements were filtered using signal processing techniques to isolate human breathing frequencies. An estimate for the breathing rate is subsequently calculated using peak detection algorithms applied to the filtered signals. The model demonstrated a proof of concept, reaching mean absolute percentage errors as low as 5.42% for breathing rate estimates. However, further analysis is needed to establish this method as a reliable estimator of breathing rate. ... Read more

The Densor is an intraoral sensor platform created to capture unique data from inside the human mouth. This thesis studies the possibility of using Densor-recorded sensor data for Voice Activity Detection (VAD) and basic non-acoustic speech (keyword) recognition. It is part of a broader effort to explore the different use cases of the Densor. This thesis summarizes the performance of existing non-acoustic wearable speech recognition devices, describes the available Densor data, details feature extraction and selection, and shows how Hidden Markov Models can beusedfor both VADandkeyword recognition. The results are promising, with an F1-Score of up to 0.73 for VAD and up to 75% accuracy for keyword recognition. While these results show that both tasks are possible, they cannot be generalized due to the small and unvaried dataset. Future work suggestions include expanding and increasing the variety of the dataset and exploring alternative models such as Conditional Random Fields and Recurrent Neural Networks. ... Read more

Detecting cancer at an initial stage could change the course of the disease's development. A non-invasive examination consists of the liquid biopsy of blood, revealing biomarkers that could provide information about the existence of a tumour or not in the organism. The research touches upon the relevance of DNA fragments, precisely the length of fragments, in the detection of cancer. An in-depth interpretation of the fragment length distribution for predicting the state of a patient as being healthy or sick with cancer was approached. The distribution was explored from four perspectives: the complete fragment length distribution, the size range from 90 to 150 bp, important lengths selected by the feature extraction methods and the Fourier Transform of the initial data. These were input in three machine learning models. Using the fragment lengths between 93 and 98 produced accuracy and AUC scores of over 0.85 for all supervised classification models. Processing the data with the Fourier Transform and using the amplitude of spectrums as features in the Random Forest model resulted in an AUC of 0.99. ... Read more

Cancer represents a huge challenge in the medical world, necessitating early detection methods to improve treatment outcomes. The field of fragmentomics emerged as a promising option towards developing efficient non-invasive cancer diagnosis tools. By analysing the differences between the cfDNA fragments from blood samples of healthy patients and patients with cancer, this study aims to determine the most important fragmentomics features for cancer detection. The methods present in this work involve extracting features from the cfDNA fragments available in the experimental dataset, applying a pipeline of feature selection techniques that removes the redundant features, training and evaluating a logistic regression and random forest classifiers to differentiate between healthy and diseased samples, and finally extracting the feature weights from the trained models to understand which features contributed the most to the classification task. Filter-based variance thresholding and Correlation-based Feature Selection (CFS) were employed to refine the dataset. Independent t-test and the Mann-Whitney U test are used to calculate the relationship between the cancer and healthy samples. The Pearson correlation coefficient calculates the correlation between each pair of features. The classification performance of the two proposed models is assessed using the train/test split and the nested cross-validation techniques. The evaluation reveals that logistic regression constantly outperforms the random forest and that removing the redundant features increases the performance of both classifiers. Certain genomic bins, mostly on chromosomes 1, 7 and 8, contain significant features for the classification task. These findings suggest that understanding the importance of the fragmentomics features can lead to improved diagnostic tools such as cancer detection based on blood tests. ... Read more

In recent years, a new way of cancer diagnostics has emerged, the analysis of DNA fragments circulating in the blood of cancer patients known as fragmentomics. This DNA, known as cell-free DNA (cfDNA), is an easily available biomarker for cell types. Deducing the tissue origin of cfDNA can reveal anomalies in cell death caused by diseases. That holds great potential in cancer detection and monitoring. The process of establishing the cell composition of a blood sample is called cell deconvolution. This research paper focuses on the comparison of two methods of cell deconvolution. The first one UXM, solves this problem by employing a reference-based technique using a methylation atlas. The second one reference-free cfSort, utilizes a Deep Learning Neural Network to perform the sample analysis. In the paper, however, a simpler architecture was trained due to the difficulties in reproduction. Experiments have been conducted to assess the sensitivity of both methods. Experiments consisted of 5 major cell types together mixed with white blood cell DNA fragments to assess the sensitivity of each method. Furthermore, different metrics such as Pearson's correlation coefficient have been used to determine the accuracy of both methods. In the end, UXM outperformed cfSort in most metrics, including Pearson's correlation coefficient, indicating its superior accuracy in deconvolution tasks. However, cfSort showed potential for higher prediction accuracy with further development and better documentation. The findings highlight the strengths and limitations of both methods. This study suggests that while UXM is currently more reliable, future improvements in cfSort could make it a viable alternative. Continued research is recommended to enhance the accuracy and transparency of these methods, ensuring their effectiveness in real-world healthcare applications. ... Read more

Recent research has indicated attributes of cell-free DNA (cfDNA) called fragmentomics
as a promising method for late stage cancer detection in a non-invasive manner. The pri-
mary objective of this research is to uncover hidden patterns and interactions that could
enhance the accuracy and sensitivity of blood-based cancer diagnostics. This study explores
the complementarity between three fragmentomics features; fragment length distribution,
and nucleotide fragment end sequence diversity and nucleosome positioning for four dif-
ferent sample groups; breast cancer, colorectal cancer, lung cancer and healthy controls.
Various machine learning techniques such as linear regression were employed to quantify
any complementary relationships between the features ... Read more

The explosive growth in data collection driven by the proliferation of interconnected devices necessitates novel approaches to data processing. Traditional centralised data processing methods are increasingly inadequate due to the sheer volume of data generated. Distributed Tiny Learning (DTL) offers a compelling solution by distributing machine learning tasks across multiple edge devices and processing data locally, thus minimising the need for data transmission to central servers. This approach is particularly beneficial in scenarios with limited network bandwidth and stringent privacy requirements, enhancing data security and compliance with privacy regulations. The advent of 6G networks, with their promise of unprecedented speed, capacity, and reliability, can further amplify the power of DTL. By providing higher bandwidth and lower latency, 6G enables more efficient data processing and communication among edge devices, thereby enhancing the overall performance and scalability of DTL systems. This integration supports real-time decision-making for applications such as autonomous vehicles, smart cities, and healthcare monitoring.

This paper conducts a comprehensive survey of the state-of-the-art in DTL, categorising scientific literature, mapping the ecosystem and tools, and addressing performance, efficiency, and scalability challenges in ultra-low-power devices within a 6G context. Additionally, it implements and benchmarks two DTL algorithms, providing practical insights into their effectiveness and operational viability. ... Read more

Spectrum sensing is a vital technology for alleviating pressure on the radio spectrum and will become more sophisticated as billions more devices come online. In the future, more advanced techniques utilizing deep learning will sense which parts of the spectrum are available to communicate on. This is a heavily researched area, but few papers demonstrate methods of deploying deep learning on the resource-constrained edge devices that will ultimately use them. One approach called Spectrum Painting augments spectrograms and detects which signals are present with a Convolutional Neural Network. We optimize this method for microcontrollers by simplifying the computation of spectrograms and using TinyML techniques. This results in over 90% accuracy on signals with a high signal-to-noise ratio and a latency of 159 ms on a 64 MHz CPU. Our findings conclude microcontrollers are capable of utilizing deep learning for spectrum sensing, but custom hardware will still be required in a real-world deployment. ... Read more

Visible light positioning (VLP) systems enable indoor positioning through a deployment of light-emitting diodes (LEDs) as transmitters and photodiodes (PDs) as receivers. A promising approach in VLP involves recording the received signal strength (RSS) to construct fingerprint samples for later use in positioning. However, achieving high accuracy demands a labor-intensive data collection process.

In this study, we propose improvements to a data cleaning and augmentation pipeline. Our improvements focus on preserving more source data during cleaning and data-based LED position estimation for more reliable data augmentation. Experimental results show that our approach maintains comparable positioning accuracy while reducing data collection efforts by over 99%. Furthermore, we conduct experiments to investigate the impact of spatially irregular data collection strategies on positioning accuracy. Finally, we deploy a machine learning model on a microcontroller to demonstrate the practical feasibility of our proposed methods. ... Read more

Nowadays, the popularity of machine learning and artificial intelligence algorithms is very high. A new research direction has emerged where the machine learning algorithms are executed on resource-constrained embedded devices. With the development of the Internet of Things paradigm, these edge devices are deployed in a lot of places. Due to the limited resources of embedded devices, it is difficult to bring machine learning algorithms to them. This is where the on-device split inference comes in. It is possible to distribute the inference between multiple edge devices and the cloud so that the edge devices can execute the inference of complex machine learning models. This paper presents a systematized literature review of papers that focus on on-device split inference for edge devices. The papers are analyzed and compared based on pre-determined questions and displayed based on several categories. ... Read more

The versatility of the internet enables many applications that play an increasingly bigger role in our society. However, users have little control over the route that their internet traffic takes, which prevents them from controlling who sees their packets and how their traffic is handled. Researchers have proposed an extension to the internet, called the responsible internet, that aims to provide users with control over the route that their internet traffic takes.
Providing this control is the aim of this thesis. Users can control their route by specifying requirements that their route has to fulfill. This thesis defines the Maximum Path Requirement Intersection (MPRI) problem as the problem of finding the route that satisfies as many of the user’s requirements as possible, and this thesis proves that MPRI is NP-hard. Subsequently, both a heuristic to solve the problem in a reasonable amount of time as well as an exact algorithm that guarantees to find the globally best path are introduced. The performance of the heuristic is measured relative to the globally optimal solution given by the exact algorithm. Results show that less features allow the heuristic to have a larger search space, which improves the results; that the runtime of the heuristic scales polynomially in the number of hops between the start and end node; that the heuristic is most effective in graphs that have a power-law degree distribution and least effective in grid-like graphs; and that in a realistic setting the heuristic runs quickly while performing close to optimal. ... Read more

Byzantine consensus protocols are designed to build resilient systems to achieve consensus under Byzantine settings, maintaining safety guarantees under any network synchrony model and providing liveness in partially or fully synchronous networks.
However, several Byzantine consensus protocols have been shown to violate liveness properties under certain scenarios. Existing testing methods for checking the liveness of consensus protocols check for time-bounded liveness violations, which generate a large number of false positives.

In this thesis, for the first time, we check the liveness of Byzantine consensus protocols by the temperature and lasso detection methods and precisely detail ad-hoc system state abstractions that should be used to test these algorithms.
We provide a theoretical analysis of the recently-published safety and liveness attacks and whether the existing test method is able to detect them.
To investigate the issues, we focus on the streamlined blockchain consensus, particularly the HotStuff protocol family, which has been recently developed for blockchain consensus. Among these protocols, the HotStuff protocol itself is both safe and live under the partial synchrony assumption, whereas 2-Phase Hotstuff and Sync HotStuff protocols can violate liveness in subtle fault scenarios.
We implemented our liveness checking methods on top of the Twins automated unit test generator to test the HotStuff protocol family and explored the scenarios with message delay.
Our results indicate that our methods successfully detect all known liveness violations and produce fewer false positives than the bounded liveness checks. ... Read more

Consensus algorithms, as well as distributed systems in general, are vulnerable to concurrency bugs due to non-determinism. Such bugs are hard to detect since it is necessary to test using a lot of different scenarios and even then, there is no guarantee to find one.

Controlled concurrency testing is a proposed solution to that problem. Its main strength is allowing for more control over the order in which threads are executed, using controlled scheduling.

In this paper, we utilize Coyote, a concurrency testing framework, to test for concurrency bugs in PBFT, the seminal consensus algorithm. Random Walk, Delay-Bounding, Probabilistic Concurrency Testing (PCT) and Q-Learning are the exploration strategies we performed experiments with. The goal is to find schedules that lead to concurrency bugs. We test for 2 bugs that are seeded by us into the algorithm.

A comparison of the results shows that fair exploration strategies outperform their unfair variations. Finally, we conclude that PCT and fair PCT are the best-performing strategies for the benchmarks we use. ... Read more