Machine learning models trained on sensitive financial records may incorporate private data without the data owner’s knowledge or consent. A bank that suspects a fintech company has trained a model on stolen customer records faces a structural dilemma: to prove misappropriation, the bank must disclose the dataset, which may itself constitute a privacy violation under Article 5 of the General Data Protection Regulation (GDPR). Existing verification techniques do not fully resolve this issue. Membership inference attacks provide probabilistic rather than deterministic evidence, dataset watermarks require proactive embedding and can be removed, and zero-knowledge proofs of training verify training correctness but do not determine whether the training data overlaps with private records.

This thesis proposes a three-party cryptographic protocol that simultaneously certifies training correctness and characterizes the intersection between the trainer’s dataset and the data owner’s private dataset, without revealing raw features or labels to the auditor. Before training, the bank posts a cryptographic commitment to its authorized records on a public ledger. The trainer posts a commitment to the training dataset and submits a proof after training. The auditor verifies two properties using only the proof and the public commitments.

For training correctness, a sumcheck protocol with multilinear KZG oracles is used, adapted from deep neural networks to linear and logistic regression, to verify that each gradient descent step was computed correctly over the committed data. For data authorization, each training sample is checked against the bank’s authorized dataset using the zero-knowledge sets construction of Silvio Micali, Michael Rabin, and Joe Kilian. These two components integrate without requiring an additional circuit. The per-sample commitment used as a lookup key is already a public element of the training commitment, allowing the auditor to derive it locally.

The zero-knowledge property of the combined protocol is formally established through simulator-based composition, providing a reusable proof structure for combining verifiable computation schemes with cryptographic set-membership proofs. Additionally, a soundness issue in the linear regression verifier was identified and corrected by enforcing error-label consistency, reducing the residual error probability to a negligible level relative to field size.

The protocol was benchmarked on the UCI Default of Credit Card Clients dataset using batch sizes up to 256 and up to 23 features. Proof size grows logarithmically with batch size and linearly with feature count, reaching approximately 80 KB per gradient step at the maximum configuration. Prover time scales linearly with dataset size and feature count, reaching approximately 3 seconds per step. The zero-knowledge membership check adds a constant 20 KB and 4.5 ms per sample, independent of dataset size. Verifier time grows linearly only with feature count, reaching approximately 500 ms per step. The computational burden therefore remains primarily with the trainer rather than the auditor.

A remaining open problem is the Consistent Data Commitment (CDC) assumption, which states that the column-wise multilinear KZG commitments and row-wise univariate KZG commitments bind identical underlying data. The current protocol does not cryptographically enforce this consistency. Developing a cross-commitment scheme to close this gap remains the highest-priority direction for future work.
... Read more

Sequential regression has many real-world applications, including energy consumption forecasting and traffic prediction. In this paper, the use of extended regression automata for sequential regression is investigated. The automata are learned using an MSE-based approach with the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to score merge and split operations, where the use of BIC is a novel contribution of this work. The approach is evaluated on three datasets and compared against persistence, a regression tree, and a random forest. The automata perform comparably on the small dataset, but on the larger datasets, they do not outperform the regression tree and random forest baselines. This is mainly caused by a bias towards splitting during learning, which prevents the automata from making use of loops and causes them to resemble trees. We further found that BIC produced smaller automata and ran faster, but its accuracy was mixed relative to AIC for our datasets. ... Read more

The goal of DFA learning is to infer a target DFA from a set of positive and negative observations. Active learning methods allow a learner to hypothesize about the target DFA by querying a teacher. Constructing a teacher that can efficiently answer these queries can significantly speed up the learning process. A distinguish query allows the learner to provide two words for the teacher and the teacher responds with a distinguishing suffix (i.e. a suffix such that if it is concatenated to the two provided words, one of the newly formed words will be among the positive observations, and the other among the negatives).
In this thesis, our aim is to design different implementations of a data structure that supports distinguish queries (DistinguishQueryDataStructure), and compare them through theoretical and experimental analysis.
To this end, we introduce the double trie, and show that using this data structure, it is possible to reduce distinguish queries to set intersection queries.
Then we design three different implementations of DistinguishQueryDataStructure, one of which uses a trie based approach with depth-first search, and two of which are based on the double trie.
Our experimental analysis shows that a double trie-based method can compete with, and even outperform in some cases the trie and depth-first-search-based approach. ... Read more

Anomaly detection is very important in this day and age. Finding anomalies in
software can help fix bugs or prevent malicious intent from third parties. One way to detect these anomalies is by making use of pdfa’s. Pdfa learning an extension of dfa learning can be used to construct a model of an application given some logs of the application. To build a pdfa, a probabilistic deterministic finite automata, various algorithms in combination with extra parameters can be used. In this paper we will use the FlexFringe framework to look into these so called search orders to answer the question: What are good search orders in extended state machine learning for anomaly detection? We looked into the Alergia and RTI+ algorithm on data with and without time information. Our main conclusions are that Alergia together with the blueblue parameter works best for creating a model for anomaly detection. ... Read more

Inferring a minimal Deterministic Finite Automaton (DFA) from a set of labelled traces is an NP-complete problem, making heuristic search strategies necessary. The red-blue framework addresses this inference process by sequentially applying merge and extend refinements to a prefix tree acceptor, with the ordering of these refinements determining the compactness of the final automaton. This paper investigates the use of Monte Carlo Tree Search (MCTS) as a heuristic for finding higher-quality refinement orderings. The problem is first formalised as a Markov Decision Process, enabling MCTS to serve as a solver. We evaluate four action-selection policies and two reward metrics across several datasets from the StaMiNa competition. Results show that the uniform-merge-first, weighted-merge-first and weighted-minimal policies outperform the uniform policy, although none of these policies dominates the others. The weighted-minimal sub-policy produces a more compact automaton than the weighted-average sub-policy, confirming that favouring merge over extend refinements during simulation is beneficial. Two reward metrics are compared: the final simulated DFA size and the number of rollout steps. Both perform similarly overall; however, the rollout steps metric has the structural advantage of being immune to reward stagnation, since the reward improves monotonically as the search tree grows. Correlation analysis further reveals that the EDSM scoring function provides no directional signal towards smaller DFAs. ... Read more

With the rise of memory costs and the persistent under-utilization of memory in clusters, researchers have begun exploring alternative approaches to improve memory efficiency and reduce operational costs. Resource disaggregation is becoming increasingly common and sought after, driven by the emergence of new interconnect standards such as CXL and, previously, OpenCAPI. While the industry is primarily moving toward memory pooling, where memory is dynamically provisioned among applications or virtual machines, this work investigates distributed memory disaggregation and sharing. IBM's Power10 processors include hardware support that enables multiple systems to directly share memory. However, few applications have been developed to take advantage of disaggregated shared memory.

Since Memory Inception, Power10 processors' memory disaggregation hardware, is not yet fully operational, a ThymesisFlow prototype, upgraded to support a shared disaggregated memory system with the help of Apache Arrow, is used to implement a practical application. The selected application is a graph processor capable of detecting money laundering patterns in financial transaction graphs in real-time. These patterns yield transaction features that machine learning algorithms can use to identify fraudulent financial transactions.

Our proof-of-concept implementation enables the creation of a distributed graph, represented as Apache Arrow tables, that can process large datasets in real-time. The graph resides in a shared disaggregated memory region and can be accessed by multiple systems without data copying, incurring lower latency penalties than network-based data retrieval. The distributed graph processor was developed and tested using the ThymesisFlow prototype provided by the Hasso Plattner Institute. ... Read more

This thesis explores the application of ahead-of-time parser generation to improve the throughput of big data ingestion. To investigate parser generation, this work produced several libraries related to the conversion of CSV to Apache Arrow.

First, an ahead-of-time parser generator was developed in the form of a Rust derive macro and a supporting library. Using a derive macro, a schema can be defined by a Rust struct definition for which a CSV to Arrow reader is derived. With the knowledge of the schema at compile-time, extra optimizations were possible. For instance, when the types in a schema are known to be of fixed size, estimates for the number of records in an input buffer could be made. With this principle, input bound checks could be reduced.

Experiments showed that ahead-of-time generated parsers outperformed state-of-the-art frameworks such as Apache Arrow, Polars, and DuckDB. Benchmarks for single types revealed that for integers, unbuffered and buffered generated parsers achieved a throughput of at least 1.5x compared to Apache Arrow, with the reduction of size bound checks sometimes even reaching a throughput of 3x. For floating point numbers, the buffered parser performed slightly better than Arrow. However, the unbuffered, and buffered with reduced size bound checking, parser achieved a throughput of at least 1.5x. For strings, the buffered parser performed similar to Apache Arrow since it uses the same efficient buffered string parsing. The unbuffered parser only achieved half the performance. Furthermore, benchmarks for parsing the TPC-H and TPC-DS datasets showed that the buffered parser generated ahead-of-time performed better for real-world datasets compared to the frameworks mentioned above. The unbuffered parser was only able to achieve a higher throughput for datasets larger than 100 MB. For the datasets, the throughput varied based on the distribution of types.

Additionally, this work explored, but did not integrate, the use of multi-threaded CSV parsing. However, experiments revealed that the performance of parallelization depends on how fast CSV can be scanned for record positions whilst correctly checking character escaping. An experiment with a custom multi-threaded parser implementation showed that when scaling the number of parse threads, the throughput is limited by scanning rather than parsing. This characteristic was also found for Polars and DuckDB, which support multi-threading. The scanning was shown to be possibly improved by using SIMD, which allowed scanning record delimiters at 2.8 GB/s using AVX2. This is approximately 1.5x more than the maximum throughput that Polars or DuckDB achieved.

https://github.com/sstreef/csv-to-arrow
... Read more

Accurate residential property valuation is essential for mortgage lending, taxation, and urban planning, yet remains challenging due to complex spatial and temporal dynamics. Traditional econometric models are interpretable but rely on restrictive assumptions, while machine learning–based automated valuation models (AVMs) improve predictive accuracy but treat transactions as independent, overlooking spatial spillovers and evolving trends. Recent graph-based approaches partially address spatial autocorrelation, but often rely on static or unidirectional structures that limit their expressiveness.

We introduce a Multi-Scale Bidirectional Spatio-Temporal Graph Neural Network (MBSTGNN) that models transactions and neighbourhoods as dynamic graphs linked through bidirectional message passing. A temporal memory mechanism maintains consistency across time, enabling the model to capture evolving market conditions. Evaluated on Rotterdam housing transactions, MBSTGNN outperforms strong baselines, particularly in sparse-data settings, and produces embeddings that reveal domain-consistent socio-spatial and temporal patterns. These results demonstrate its potential for advancing automated valuation and related spatio-temporal prediction tasks. ... Read more

Heterogeneous datasets hold a large percentage of all digital data that is available. With the rise of the digital medium, they have played a strong part in addressing the need for a structured way of storing data, particularly through the use of relational databases. To better leverage such data, the consensus of researchers has been in favour of using Graph Neural Networks to make predictions and infer pos- sible outcomes. With the rise of the Transformer model and the clear limitations that GNNs inherently have due to their over-smoothing and over-squashing properties, a clear transition occurred into combining and leveraging the properties of both GNNs and Transformer models, Graph Transformers. While this method has been better researched in the context of homogeneous datasets, much remains unexplored in its heterogeneous counterpart. This paper tackles this by applying Graph Transformers to multiple heterogeneous datasets, examining the differences and advantages between interleaved and cascade architectures of GTs and how the homogeneous positional encodings transfer to the heterogeneous context. ... Read more

The growing volume of online misinformation has increased the demand for automated fact-checking systems. While large language models (LLMs) have demonstrated potential in this domain, real-world claims are complex that challenge standard prompting techniques. These claims are multi-aspect, requiring integration of evidence across different sources. Among these, a challenging subset are claims with Conflicting labels, where different parts of the claim support opposing veracity labels. Both cases demand nuanced, context-sensitive reasoning that LLMs struggle to perform reliably. This thesis investigates two methods to enhance LLM
reasoning for fact-checking - claim decomposition and test-time scaling. Claim decomposition
breaks down complex claims into simpler sub-questions, promoting more structured reasoning. While this improves performance on Conflicting claims, it can degrade accuracy for straightforward claims, particularly those labeled as True. To mitigate this, an adaptive decomposition strategy is proposed, selectively applying decomposition only when beneficial. A taxonomy of reasoning failure - termed overthinking - is identified, where the model becomes unnecessarily strict due to noisy evidence or overly specific sub-questions. To further address this issue, test-time scaling using a reward model is employed to rank candidate outputs.
This approach yields an 18.8% relative improvement in macro F1-score over the baseline and reduces overthinking by encouraging context-aware leniency. Together, these findings underscore the importance of targeted reasoning strategies for improving the robustness and reliability of LLM-based fact-checking. ... Read more

Graph Neural Networks (GNNs) have achieved state-of-the-art performance in various applications due to their ability to capture complex structural relationships within graph data. However, their inherent black-box nature poses significant challenges to model interpretability, particularly in the context of link prediction tasks. This thesis investigates the explainability of GNNs for link prediction, a critical task in domains such as social network analysis and recommendation systems. We propose an efficient extension of the Zorro explainer, originally designed for node classification, to address the unique challenges of explaining link predictions. Our method leverages top similar nodes for initialization and incorporates techniques to accelerate the greedy search process, ultimately producing binary node and feature masks as explanations. Additionally, we introduce a novel Edge-level RDT-Fidelity metric to better evaluate explainers that generate edge masks, complementing existing metrics like RDT-Fidelity and Sparsity. Experimental results on the Cora and PubMed datasets demonstrate the superior performance and efficiency of our Zorro extension. This work contributes to the growing field of GNN interpretability by providing new methodologies and evaluation frameworks specifically tailored for link prediction. ... Read more

This thesis addresses the growing demand for faster and more reliable data transfer solutions in data-intensive applications, with a strong emphasis on scalability and flexibility. The research centers on the design and implementation of a high-speed data transfer system that utilises PCIe over fiber optics. The main objective is to explore how PCIe technology can be effectively extended over long distances using fiber optics, while maintaining high throughput and low latency, essential for supporting the demands of modern data-intensive environments.

The design and implementation of the system were approached methodically, with careful selection of components and a focus on optimising the system architecture for performance. Detailed experimentation was conducted to evaluate the system's capabilities, resulting in successful data transfer rates of up to 58.02 Gbps using a Gen3 x8 PCIe configuration over a 10-meter fiber link. These findings confirm the viability of PCIe over fiber as an effective solution for high-speed data transfer, while also identifying specific areas where further improvements could be made.

Future work could explore the adoption of newer PCIe generations and the optimisation of critical parameters, such as Maximum Payload Size (MPS). Additionally, there is potential to enhance the system's performance by investigating optical connectors capable of supporting sideband signals. These advancements would further improve the system's adaptability and efficiency across a wide range of application scenarios.
... Read more

Port-Scanning is a popular technique that helps detect open ports to connect to on the internet, with both benign and malicious applications. While methods have been developed to detect scans coming from one source, adversaries have started to distribute their scans among multiple hosts. Detecting these collaborative scanners, however, is a difficult task, and no established method has emerged yet. In this paper, we propose a method to cluster scanning groups from network telescope data based on the assumption that scanners working together will have similar behavioral features. An evaluation framework based on address space coverage and overlap is introduced and the performance of two clustering algorithms, HDBSCAN and DBSCAN, is compared. A clear indication was found that only relying on behavioral features and applying unsupervised clustering techniques can result in incomplete groups or clusters with more than one group inside, with HDBSCAN generally performing better than DBSCAN. The effectiveness of the evaluation framework is also discussed, as some of the identified groups indicated overlaps, but manual analysis reveals that they have the potential to belong to a single party. ... Read more

Detecting distributed scans is crucial for understanding network security threats. This research uses an algorithmic approach to identify collaborative ZMap scanning activities in the network telescope data from TU Delft. ZMap is a high-speed network scanner capable of scanning the entire IPv4 address space. The main research question centers on creating an algorithm for detecting these distributed scans. The research method includes analyzing network telescope data, examining ZMap packets and modifying the set cover algorithm to detect collaborative ZMap scans. Key contributions include adapting the set cover algorithm to find sources that perfectly cover the entire destination address range without overlaps, which is a unique feature of ZMap scans. Results indicate that this method effectively identifies coordinated scanning activities. It is concluded that utilizing network telescope data and an adapted version of the greedy set cover algorithm significantly improves the detection of distributed scanning operations using ZMap. ... Read more

The first step of many cyber attacks is the reconnaissance phase. One of many reconnaissance methods employed by adversaries is internet-wide scanning, which
probes the entire internet to find which hosts have open ports. These scans are practically
impossible to detect by a firewall or Intrusion Detection System if an attacker chooses to
distribute their scan on multiple hosts. Many of these scans embed a fingerprint in their
packets, which can easily be detected if they are known. Previous studies have developed
an algorithm that is able to identify these fingerprints, but they were not able to identify
fingerprints for large portion of their data. This study proposes an iterative approach
using stratified sampling, in order to see how this affects accuracy. An experiment showed
the algorithm is able to identify fingerprints for sets of packets that make up less than
0.5% of all packets, and less than 0.0001% of sources. Analysis of the fingerprinted groups
indicated that these groups are not part of a collaborative scanner, but hold for the same
fingerprint by coincidence. ... Read more

This research investigates the effectiveness of combining Feature Tokenizer Transformer (FTTransformer)[6] with graph neural networks for anti-money laundering (AML) applications. We explore various fine-tuning techniques, including LoRA[9] and vanilla fine-tuning, our baseline FTT architecture. Using the IBM AML dataset [1], we compare the performance of different models and fine-tuning approaches. Our results indicate that FTT alone do not outperform GNN’s and careful configuration is required when working with datasets of Multi-Modality. This work contributes to the development of more efficient and accurate methods for detecting financial fraud patterns. ... Read more

Port scanning is a technique often used by adversaries to detect vulnerable services running on a machine. There are defense mechanisms in place that can detect fast, single-source port scanning, but one of the ways to remain hidden is to distribute the scan between multiple hosts. These distributed groups of machines can divide the address space and collaboratively scan the whole Internet within minutes and remain relatively hidden.
This paper proposes a simple method to detect these collaborative scanners based on the TCP/IP header and demonstrates its efficiency. It also tracks these scanners for a longer period and describes their behavior and how they develop over time. This includes the infrastructure they utilize, the specific ports they target, and additional relevant details. This perspective has not been previously explored in the academic literature and we find it to be important such that defenders get a better understanding of the threats they are facing. ... Read more

This paper investigates the effectiveness of various clustering algorithms in detecting collaborative Internet scanning groups. The packet dataset used is collected from TU Delft's network telescope, and is aggregated into scanning sessions and analyzed using K-Means, Hierarchical Clustering, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Clustering Using Representatives (CURE), and Bradley-Fayyad-Reina (BFR). This paper also introduces an evaluation framework based on five degrees of certainty to assess the likelihood that a cluster is collaboratively scanning. The findings indicate that DBSCAN consistently outperforms other methods in identifying collaborative scanning groups, while CURE shows superior performance to BFR, K-Means, and Hierarchical Clustering. It is hoped that these insights help provide a strong foundation for enhancing network security through improved detection of collaborative scanning behaviors. ... Read more

Passage re-ranking is a fundamental problem in information retrieval, which deals with reordering a small set of passages based on their relevancy to a query. It is a crucial component in various web information systems, such as search engines or question-answering systems. Modern approaches for building re-ranking systems rely on neural language models such as BERT, or its derivatives, to create dense indexes for the target document corpus. While such approaches bring significant performance gains compared to classical lexical re-rankers, they have the disadvantage of increased memory costs.

A family of methods that can be used to reduce the memory footprint of a dense index is called vector quantization. Vector quantization algorithms usually rely on a combination of clustering and space manipulation operations to perform a lossy compression of the dense index at the expense of index performance. While vector quantization is widely used for first-stage retrieval, its use in the context of re-ranking is underexplored. To this end, this thesis evaluates the effectiveness of product quantization, a well-known vector quantization method, on single-vector dual-encoders, specifically TCT-ColBERT and Aggretriever. In addition to this, we show how linear interpolation of sparse scores can be leveraged to improve the performance of quantized dense indices with negligible costs to the memory footprint or speed. Last but not least, we propose WolfPQ, a learnable quantization method aimed at further improving quantization for re-ranking by bridging the gap between the objective functions used in training product quantization and re-ranking systems, respectively. ... Read more