Graphs are everywhere in the real world. Although Graph Neural Networks (GNNs) have demonstrated strong performance across a wide range of graph-based tasks, their unique neighbourhood aggregation mechanism encodes sensitive feature, label, and topological information into model weights, making them vulnerable to Membership Inference Attacks (MIA). This poses challenges for organizations regarding the practical deployment of GNN models trained on their private data, as they have the obligation to protect the privacy of individuals in the dataset. While numerous privacy-preserving techniques, including differential privacy (DP) based approaches such as PrivGNN, have been proposed, they are subject to the inherent utility-privacy trade-off where achieving strong privacy often comes at the cost of reduction in model utility. In this work, we investigate Self-Supervised Learning (SSL) pretraining on a public graph, with a focus on contrastive pretraining via GRACE, as a utility-preserving privacy mitigation strategy with the potential to improve privacy without sacrificing generalizability. All experiments are performed under an inductive setting with disjoint graphs, where the model is evaluated on nodes never seen during training. We first show that SSL improves the utility-privacy trade-off in standard GNN training, with the backbone learning rate (BLR) acting as a tunable knob to navigate the trade-off. Subsequently, we perform a node-level vulnerability analysis, finding that structural isolation and intra-class feature dissimilarity act as predictors for nodes' privacy vulnerability, while neighbourhood class divergence exhibits more dataset-dependent results. We then show that employing lower BLR alone can disproportionately reduce the privacy vulnerability of the most exposed nodes, with SSL providing additional, uniform privacy benefit on top of it. Beyond standard GNN training, we integrate SSL into DP-based PrivGNN and show that it substantially recovers the utility lost to privacy mechanisms, providing significantly tighter formal privacy guarantees without hurting empirical privacy. ... Read more

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

Software correctness is a hard problem. Dependently-typed programming languages like Agda help us to provide guarantees about software while writing it in a Correct-by-Construction (CbC) manner. However, the implementation of a dependently-typed programming language may contain bugs itself: Agda has bugs in its own implementation on a regular basis. Dependently typed-typed programming languages should therefore verify themselves in a Correct-by-Construction manner using their own components: self-verification. Agda Core aims to do this for Agda: it is a core language for Agda and provides a Correct-by-Construction type checker for itself derived from a trusted type theory, written in Agda.

Before Agda Core can be used as a true self-verifier for Agda and be integrated into Agda’s main compiler and type checker, more features from Agda need to be supported by it. In this thesis, we focus on Agda’s η-conversion for function types and record types. We formalize η-conversion for function types and record types with at least one field using untyped conversion: we also add support for records to Agda Core along the way. We also show progress towards a formalization of Agda’s η-conversion using typed conversion, which allows for formalizing the often tricky-considered η-conversion for Agda’s unit type. For all of these formalizations, we show that they can be added to Agda Core with reasonable effort. Overall, this work therefore provides an important step towards the ultimate goal of a self-verified type checker for Agda with support for all of Agda’s features, which decreases the future potential for bugs in Agda’s implementation with relation to η-conversion, and teaches us how to self-verify η-conversion for a dependently-typed language. ... Read more

Post-quantum cryptography has emerged in response to the increasing threat posed by quantum computers to classical public-key systems. Subject to the ’store now, decrypt later’ strategy, data is already vulnerable. In response, the National Institute of Standards and Technology (NIST) standardized ML-KEM, a lattice-based key encapsulation mechanism. Because this algorithm is more computationally intensive than previous methods, acceleration is required for practical application. One of the bottlenecks, the Number-Theoretic Transform (NTT), relies heavily on modular multiplication. The novel Plantard modular multiplication method has been shown to improve performance by up to 12% in software by reducing the number of multiplications per modular multiplication, but it has not been examined in hardware for ML-KEM. Therefore, in this thesis, a design and evaluation of a Loosely-Coupled Accelerator (LCA) containing a Plantard modular multiplier executing the ML-KEM.KeyGen function on an FPGA is presented. The design is parameterized, resulting in configurations that use the Plantard or traditional Montgomery multiplier on either 1 or 2 parallel Unified Butterfly (UB) modules to execute the NTT, which are evaluated on the resulting Area-Time Product (ATP). The results show that, in a pipelined design, Plantard’s reduction of 3 to 2 multiplications per modular multiplication results in a marginal 1-clock-cycle reduction per applied transformation. This minor improvement of less than 0.03% of ML-KEM.KeyGen’s performance is outweighed by the additional resources Plantard requires to support wider operands and its increased ROM utilization for storing larger constants. This results in a 2-UB configuration using Montgomery multiplication that achieves the best ATP, executing the ML-KEM-1024.KeyGen function in 38k and the NTT in 580 clock cycles, respectively, at 100 MHz. These findings suggest that Plantard multiplication is better suited to resource-constrained or non-pipelined applications. ... Read more

Automated Static Analysis Tools (ASATs) generate a massive volume of non-actionable warnings. To address this, this thesis investigates the performance and resource trade-offs between classical Machine Learning (ML) models and Large Language Models (LLMs) for generating actionability probability scores. Utilizing the NASCAR dataset of over 1.2 million Java warnings, we evaluate optimized classical models (Random Forest and Logistic Regression) against the Claude 4.x LLM family using classification metrics (F1-score, AUC) and probabilistic calibration (Brier scores), supplemented by a qualitative user study of 15 industry professionals. Empirical results demonstrate that an optimized Random Forest yields superior predictive performance (F1-score: 76.85\%, AUC: 0.87) and reliable uncertainty calibration (Brier score: 0.1549), rendering the massive computational overhead of miscalibrated LLMs unnecessary. However, the user study identifies a human-AI feature disconnect: while the Random Forest relies heavily on historical metadata, developers universally demand source code context and severity indicators. Ultimately, an optimized Random Forest provides a significantly more efficient framework for scoring ASAT warnings, provided the scores are tightly coupled with the structural evidence required to sustain human trust. ... Read more

Subgraph pattern detection aims to uncover complex interaction structures, for example those associated with money laundering in financial transaction networks. State-of-the-art graph neural network (GNN) solutions, however, assume centralized access to the entire graph. When the graph is instead distributed across multiple financial institutions, each client computes node representations using only its own subgraph, so client-local GNN computations diverge from those of a centralized model. We formalize this divergence as the structural observability problem, in which subgraph patterns crossing partition boundaries become locally unidentifiable. This divergence manifests as both a forward gap in the node representations and a backward gap in the training-time adjoint signal. To close both gaps, we propose a per-step, layer-wise exchange framework with two complementary components: a forward exchange that synchronizes node representations at every layer of the forward pass, and a backward exchange that synchronizes the corresponding gradient signals at every layer of the backward pass; neither component exposes raw features or labels. Under an extended subgraph assumption and shared model parameters across clients, we prove that the forward exchange recovers the representations a centralized GNN would compute over the full graph (representation equivalence) and the backward exchange makes the per-client parameter gradients sum to the exact centralized gradient (gradient equivalence). Together, forward and backward exchange make federated training equivalent to centralized training. Experiments on synthetic directed multigraphs with cycle, biclique, and scatter-gather patterns show that forward exchange and federated parameter aggregation are complementary rather than interchangeable, and that their combination recovers most of the gap to centralized performance. This recovery depends on per-step freshness, with stale per-epoch exchange leaving a measurable residual. Adding backward exchange yields further improvements, with the largest gains achieved when the cross-client connectivity is densest. ... Read more

Machine learning models can separate diseased from healthy patients using gut microbiome data, but the explainability techniques used to justify their predictions were developed for ordinary tabular, image, and text data, not for the compositional, high-dimensional, and sparse abundance tables produced in microbiome studies. We test how far these techniques can be trusted in this setting through a comparative study, applying multiple explainability techniques to multiple models across three levels of representation: classical models on the raw abundance table, the same models on a foundational model’s latent space, and end-to-end deep models. The task is to distinguish Inflammatory Bowel Disease from controls using the IBDMDB dataset.

We identify three failure modes that a single-model analysis would hide. First, the way an evaluation metric is measured or implemented can influence the metric as much as switching the model under test. Second, predictive accuracy should be interpreted separately from biologically correct reasoning. Third, an encoder’s prior can exert a stronger influence on the result than the model built on top of it. These findings suggest that explanations of microbiome models should be evaluated across multiple measurement methods, considered separately from accuracy, and interpreted with representation effects in mind. Otherwise, any single explanation may reflect substantial artefacts inherent to its implementation or measurement method. ... Read more

The gut microbiome is linked to many diseases, but its sparse, high-dimensional, and compositional structure makes representation learning challenging. Recent microbiome foundation models suggest that deep learning, especially transformers, may capture useful community-level patterns, but architectures are rarely compared under equivalent conditions.

This study compares an autoencoder, a variational autoencoder, a transformer-based autoencoder, and a pretrained MGM benchmark on the same genus-level gut microbiome dataset. Models are evaluated using held-out AUROC for disease-status prediction and silhouette scores for latent-space structure.

The results show that representation learning does not substantially improve prediction over directly using preprocessed abundance features. The transformer-based autoencoder achieves the highest AUROC, but only marginally outperforms the random forest baseline. The autoencoder performs similarly, while MGM with limited fine-tuning and especially the variational autoencoder perform worse. Silhouette scores are close to zero for control--disease separation and negative for disease-label and disease-category clustering, indicating weak biological structure in all embedding spaces.

Overall, this study shows that architectural complexity alone does not guarantee better microbiome representations. ... Read more

Microbiome machine-learning models often perform well within one dataset, yet their ability to generalize across studies and disease contexts remains unclear. This is difficult because microbiome data are heterogeneous in cohort composition, sequencing pipelines, and feature spaces.

To address this, this study benchmarked crossstudy microbiome classification under four evaluation settings: study-to-study transfer, leave-onestudy-out transfer, disease-context transfer, and within-study prediction. Random Forest was compared with TabPFN and MGM, two foundationstyle models, under multiple feature-space alignment strategies, including union, intersection, and
prevalence-based filtering.

The results show that external validation is harder than within-study prediction (average ROC-AUC: 0.56–0.66 vs. 0.75). Random Forest consistently achieved the strongest performance on the external transfer tasks, generally outperforming both TabPFN and MGM. Feature-space alignment affected performance, but its impact was modest compared with the much larger effect of the evaluation setting itself.

Overall, this study provides a reproducible benchmark for microbiome generalization and shows that robust transfer across heterogeneous cohorts remains harder than achieving high performance within a single dataset. The benchmark offers a starting point for future work on stronger transferoriented models and more realistic evaluation protocols. ... Read more

Gut microbiome data is compositional, sparse and high-dimensional, and choosing an appropriate preprocessing method before machine learning is a decision that most published studies make without systematic justification. We benchmarked 15 preprocessing methods against 8 classifiers on the Human Microbiome Project 2 (HMP2) metagenomic dataset from the Inflammatory Bowel Disease Multi'omics Database (IBDMDB), evaluating three-class inflammatory bowel disease classification (Crohn's disease, ulcerative colitis, non-IBD). Classifiers included two tree-based methods (Random Forest, XGBoost), two support vector machines (linear and radial basis function kernel), a multi-layer perceptron, elastic net logistic regression, and TabPFN.

Preprocessing sensitivity was strongly classifier-dependent. Tree-based methods were robust over all variants, with standard deviations across preprocessing below 0.01 and no significant pairwise differences after Benjamini-Hochberg correction. Scale-sensitive classifiers showed substantial effects, with the MLP exhibiting a standard deviation of 0.052 across preprocessing variants and SVM-RBF showing 73 significant pairwise differences out of 105. Distribution shape emerged as the primary driver of scale-sensitive classifier performance. Rank and quantile normalization consistently matched or exceeded theoretically motivated log-ratio transforms, while raw count-based variants such as TSS and log-TSS were actively harmful for MLP. Feature importance rankings were highly stable across preprocessing choices for tree-based models (RF mean Spearman ρ = 0.986, XGBoost ρ = 0.948), with greater variability for linear models (Elastic Net ρ = 0.772, SVM-linearρ = 0.838). This indicates that preprocessing affects scale-sensitive classifiers not only in performance but also in which features are prioritised. These results suggest that tree-based models can be used with any preprocessing choice, while scale-sensitive classifiers require careful selection of a distribution-normalising transform. ... Read more

The widespread adoption of wireless communication devices has led to increasingly congested wireless networks, creating a need for alternative communication technologies. A promising alternative is Visible Light Communication (VLC), which, instead of using radio frequencies, leverages the visible light spectrum to enable data transmission. Unfortunately, using visible light to communicate comes with its own set of limitations, such as the relatively high power consumption of sustaining a light source. In pursuit of lower power consumption, passive VLC has recently been gaining attention. Passive VLC is a form of VLC that uses ambient light to transmit data. Typically, it does so by modulating sunlight or artificial lights using Liquid Crystals (LCs). However, LCs, amongst other things, have very limited modulation speeds. This limitation has prompted the search for transmitters capable of achieving higher modulation speeds. Recent works have investigated the use of Digital Micromirror Devices (DMDs), which can achieve significantly higher modulation speeds. Unfortunately, DMDs require precise alignment for both incoming and outgoing light, imposing strict alignment constraints. To avoid these constraints, we investigate alternative transmitter designs using mechanical actuators. This has led to the design of two novel passive VLC systems dubbed DiscoLink and SpeakerLink. The DiscoLink transmitter, which makes use of a stepper motor to oscillate a multitude of small mirrors, achieves a throughput speed of 66 bits per second. Meanwhile, SpeakerLink, which makes use of a voice coil to oscillate a DVD, achieves a throughput speed of 20 bits per second. Their unique designs are enabled by using an event camera as a receiver, leveraging its high spatial and temporal resolution. Both systems operate on the principle of oscillating a reflective surface to alternate between different origins of ambient light reflected toward the receiver. Therefore, unlike other passive VLC systems, these transmitters do not have traditional ”on” and ”off” states between which they can alternate. The designs also necessitate the development of a novel modulation scheme, for which we propose two distinct methods. Although both DiscoLink and SpeakerLink do not face the alignment constraints of DMDs, they do face other challenges, such as low throughput speeds, reduced range, and noise. Whilst the practical use of these systems is limited due to these challenges, they highlight a new potential design space for passive VLC. ... Read more

Current digital payment solutions are fragile and offer less privacy than traditional cash. Their critical dependency on an online service used to perform and validate transactions makes them void if this service is unreachable. Moreover, no transaction can be executed during server malfunctions or power outages. Due to climate change, the likelihood of extreme weather increases. As extreme weather is a major cause of power outages, the frequency of power outages is expected to increase. The lack of privacy is an inherent result of their account-based design or the use of a public ledger. The critical dependency and lack of privacy can be resolved with a Central Bank Digital Currency that can be used offline. This thesis proposes a design and a first implementation for an offline-first digital euro. The protocol offers complete privacy during transactions using zero-knowledge proofs. Furthermore, transactions can be executed offline without third parties and retroactive double-spending detection is facilitated. To protect the users’ privacy, but also guard against money laundering, we have added the following privacy-guarding mechanism. The bank and trusted third parties for law enforcement must collaborate to decrypt transactions, revealing the digital pseudonym used in the transaction. Importantly, the transaction can be decrypted without decrypting prior transactions attached to the
digital euro. The protocol has a working initial implementation showcasing its usability and demonstrating functionality. ... Read more

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. ... Read more

The Dutch railway system is one of the most densely used systems worldwide and the busiest in Europe. Given the tight schedules, incidents can quickly cascade through the entire country if not handled properly. Alternative timetables are created to help train traffic controllers swiftly resolve such incidents. These schedules are currently created manually, but the team cannot keep up with demand. This is why CGI is developing VGB Solver, an application to automatically generate such timetables. The solver uses a branch and bound algorithm in which nodes are processed in best-first order, based on a heuristic value. For this thesis, different performance optimisations for this algorithm were implemented and analysed.

Various alternative formulations of the heuristic value, used to determine the likeliness of a node leading to a good solution, showed promising results.
One of these formulas resulted in solutions for 96% more scenarios than before, and improved the quality of solutions for other scenarios.

Using machine learning, a decision tree was created to predict whether applying another new formula for the heuristic value gives better results than the old formula. This classifier achieved an accuracy of 73.5% on the test data. Having the solver choose between the old and new formula based on that classification resulted in some scenarios with worse scores, but twice as many improved, and the average improvement was higher than the average deterioration.

Recommendations are made to conduct further experiments related to the heuristic value calculation. Furthermore, it is suggested to separate the scores for evaluating solution quality from the heuristic value formula, to facilitate more fine-grained changes to the calculation of the heuristic value.
... Read more

Foremost among the challenges of the Bitcoin blockchain is the scalability bottleneck. To address this issue, the Lightning Network, a payment channel network, was created. Lightning is a payment channel network that is source-routed and uses onion routing, like Tor. However, unlike Tor, the routing path is determined by optimizing a cost function, which uses public information. In this paper, a timing attack is evaluated by simulation of the Lightning Network. The goal of the attack is for an adversary that is part of a payment to determine the destination of the payment. The influence colluding adversaries have on the performance of the attack is evaluated. Three types of colluding adversaries (closest to destination, farthest from destination, average of all adversaries) are compared to a simple adversary that guesses the next hop in the payment is the destination of the payment. It is found that the adversary closest to the destination performs the best. Furthermore, shadow routing, a mitigation against these types of attacks, is evaluated against this attack. It is found that shadow routing does not have a significant impact on the performance of this attack. ... Read more

Call graphs are useful tools for representing method relationships within software projects and correlations between dependencies. Although static analysis is a prevalent method for call graph construction, it has its limitations such as struggling with handling dynamic features and lambda expressions. In this research, we introduced an approach that utilizes test suites from public Java Maven projects to construct dynamic call graphs and then merge them with static call graphs. Our objective is to explore the efficacy of the merged call graph in uncovering additional information. We employed OPAL and Soot to generate static call graphs and Java agents to trace edges during actual execution. Subsequently, a merging procedure, coupled with a filtering mechanism, was implemented to eliminate duplications. We conducted vulnerability detection analysis to assess the results and a version analysis to investigate the potential for extending our approach by merging multiple versions. Our results indicated that the merged call graph offers a modest increase in edges compared to solely static analysis. Additionally, we discovered that vulnerability identification followed a comparable pattern, supporting the consistency of our methodology. Additionally, we discover that combining several minor/patch versions of a project is a successful tactic for enhancing test coverage. Our research highlights the value of using test suites from open-source projects to build more in-depth call graphs. ... Read more

Federated Learning is highly susceptible to backdoor and targeted attacks as participants can manipulate their data and models locally without any oversight on whether they follow the correct process. There are a number of server-side defenses that mitigate the attacks by modifying or rejecting local updates submitted by clients. However, we find that bursty adversarial patterns with a high variance in the number of malicious clients can circumvent the existing defenses. We propose a client-self defense, LeadFL, that is combined with existing server-side defenses to thwart backdoor and targeted attacks. The core idea of LeadFL is a novel regularization term in local model training such that the Hessian matrix of local gradients is nullified. We provide the convergence analysis of LeadFL and its robustness guarantee in terms of certified radius. Our empirical evaluation shows that LeadFL is able to mitigate bursty adversarial patterns for both iid and non-iid data distributions. It frequently reduces the backdoor accuracy from more than 75% for state-of-the-art defenses to less than 10% while its impact on the main task accuracy is always less than for other client-side defenses. ... Read more

Machine learning models are increasingly being used within software engineering for their predictions. Research shows that these models’ performance is increasing with new research. This thesis focuses on models for method name prediction, for which the goal is to have a model that can accurately predict method names. With this thesis, we could create a tool that can suggest method names to software developers, which would assist in improving the quality of the projects.
This research aims to get insight into the robustness vulnerabilities of a method name prediction model. We use a genetic search algorithm that looks for these robustness problems. The main question this thesis tries to answer is to what extent the performance metrics are affected by applying metamorphic transformations to the test set of a trained code2vec model. Besides this, this thesis also proposes an alternative metric called percentage MRR, which might better reflect the robustness of a model. The main idea behind this metric is that it penalizes the prediction certainty of a model instead of penalizing the prediction rank.
To answer this research question, a tool is created that runs a genetic algorithm applying these metamorphic transformations to a dataset that a trained model is then evaluating. With this tool, we conducted 22 genetic search experiments on primary metrics and combinations of metrics to see the trade-offs in the Pareto fronts. The guided search of applying metamorphic transformations on the test set results in an average performance decrease of around 19%. This thesis also compares this drop in performance to the performance decrease a random search algorithm would create. Notably, for every transformer added, the average decrease in performance becomes smaller, and there are transformations, e.g., the if-false-else transformation, that have a bigger effect than others. This thesis concludes that the trained model is not robust against metamorphic transformations and has a significant performance drop. ... Read more

Searchable Symmetric Encryption (SSE) schemes provide secure search over encrypted databases while allowing admitted information leakages. Generally, the leakages can be categorized into access, search, and volume pattern. In most existing Searchable Encryption (SE) schemes, these leakages are caused by practical designs but are considered an acceptable price to achieve high search efficiency. Many attacks on SSE schemes have shown that such leakages could be easily exploited to retrieve the underlying keywords for search queries. Each attack abuses a different leakage pattern and uses different techniques to achieve high query recovery accuracy. An attacker could be passive or active, where an active attacker can inject files in an SSE scheme, while a passive attacker only observes the queried data. Some passive attacks use the number of files returned by a query to create a match with a candidate keyword. Others use the co-occurrence of multiple keywords in the files to match a query with the same occurrence. We continue this research and design a new Volume and Access Pattern Leakage-abuse Attack (VAL-Attack) that exploits both the access and volume patterns. Our proposed attack only leverages leaked documents and the keywords present in those documents as auxiliary knowledge and can effectively retrieve document and keyword matches from leaked data. Furthermore, the recovery performs with great accuracy and without false positives. We compare VALAttack with two recent well-defined attacks on several real-world datasets to highlight the effectiveness of our attack and present the performance under popular countermeasures. ... Read more

In this Thesis a fast, low-cost, anthropomorphic robotic hand with tactile feedback is designed. The hand consists of an index finger and a thumb, both of which have four degrees of freedom. All eight degrees of freedom are fully actuated using eight servomotors whose forces are transferred using tendons. The design introduces tendon routing to minimize the change in tendon tension as the joints rotate. To accomplish this at the 2 degree of freedom joints of thumb and index finger, a novel design for the ball and socket joint is created. The hand can be controlled using a tactile glove called the SenseGlove which also enables the user to receive force feedback. Testing reveals that the robot can open and close its fingers within 275 ms, and is able to grab a variety of objects. ... Read more