Core-guided solvers and Implicit Hitting Set (IHS) solvers have become ubiquitous within the field of Maximum Satisfiability (MaxSAT). While both types of solvers iteratively increase the solution cost until a satisfiable solution is found, the manner in which this relaxation is performed leads to the belief that these techniques are wholly separate from one another. This work exhibits the difficulty of directly translating the cores found by an IHS-solver to cores utilizable by an OLL-solver due to an inherent disconnect between the manner in which both approaches explore the solution space. It will then be shown how this translation can be performed more easily given certain assumptions. Finally, several techniques are introduced for performing a translation from cores of the original formula to OLL-cores which avoids the aforementioned issue, resulting in a hybrid IHS-OLL algorithm. The comparison between the performance of the hybrid algorithm and RC2 shows that the hybrid algorithm achieves analogous performance in terms of the number of instances solved while indicating that utilising cores of the original formula as a starting point for an OLL solver can be beneficial for the performance of the solver in certain cases.

Nowadays, the online industry contributes to a multi-billion dollar business, facilitates most of the population's everyday activities, and processes vast amounts of data, including personal data. Current work aims to explore the inconsistency or consistency of the content obtained by the websites to generate cookies based on various data that the user provides when visiting a web page, being by explicit consent or not. Some websites integrate with third-party companies that track users and collect their data. For this research, a custom-made Selenium-based web application (crawler) visits the Top 50 Alexa most visited websites and observes the cookies that are collected before the user's consent. After a brief data set analysis, a significant inconsistency in the cookies' count that deviates per location, device type, and operating system is detected. The results show that some websites collect private data, even though users are not informed about the collected data, and the consent for using cookies is not retrieved. These results imply that tracking persists as a serious concern. Tracking raises ethical and legal matters due to its potential adverse effect on users' data. That is why it is essential to analyse the content these trackers obtain, e.g., location, internet protocol address, and browsing history. Subsequently, to suggest possible techniques to avoid tracking.

Blockchain technology has proven to be a promising solution for decentralized systems in various industries. At the core of a blockchain system is the peer-to-peer (P2P) overlay, which facilitates communication be- tween parties in the blockchain system. Recently, there is increasing evidence that this P2P overlay plays a major role in limited transaction throughput, limited scalability, and security threats in a blockchain system. In this work, it is explained that this bottleneck is caused by the fact that currently deployed P2P overlays cannot fully capture the complexity of a blockchain system. We propose a novel approach to address these issues by introducing a novel overlay architecture. Our approach, Sovereign Domains P2P Overlays, allows for nodes in a single network to belong to multiple heterogeneous groups (called domains), each with their own set of protocols tailored to the characteristics and needs of the nodes in that domain. To demonstrate the effectiveness of the proposed overlay architecture, two novel node discovery protocols, FedKad and SovKad, were designed and implemented. The simulations conducted in a custom-built simulator show that SovKad outperforms node discovery in a Structured P2P Overlay (Kademlia) and node discovery in a Federated P2P Overlay (FedKad), providing evidence of the potential of the Sovereign Domains P2P Overlay architecture in blockchain systems.

Explainable AI (XAI) has gained increasing attention from more and more researchers with an aim to improve human interaction with AI systems. In the context of human-agent teamwork (HAT), providing explainability to the agent helps to increase shared team knowledge and belief, therefore improving overall teamwork. With various backgrounds and characteristics of humans, expert video gamers are found to have better perception and cognitive ability. This study aims to study the effect of information amount in explanations on four factors: subjective workload, teamwork performance, trust, and explanation satisfaction in different expertise levels in human-agent teamwork. To investigate the research question, we designed a simulated search and rescue task, encompassing two types of explanations: the one containing less detailed information, and the other presenting more detailed information. After conducting the experiment with 42 participants, we first divided all participants into three expertise levels based on their self-reported game frequency and the mock task score in the tutorial. Then we statistically analyzed the effect of information amount and expertise levels on the subjective workload, team performance, trust, explanation satisfaction, and activity level. In conclusion, we did not find evidence that adapting the information amount in explanations to gaming expertise levels can yield an improvement in the user experience during simulated search and rescue tasks. However, subjective workload is found to have a negative effect on explanation satisfaction. For future studies, it may be worth investigating whether expert gamers require explanations with very detailed information in HAT.

Virtual robot soccer competitions have been rising in popularity in recent years, due to their easier accessibility than physical robot soccer. This is also the reason these virtual competitions were created. The goal is to compete in the AI world cup, one of the virtual soccer competitions. A coach-based framework has already been built, and the aim is to create strong defensive capabilities for the robot team. This paper in particular focuses on covering opponents without the ball. The marking strategy and specific opponent covering that have been devised for this are explained. Both are completely rule-based, and opponent covering has been thoroughly tested with an in-house developed batch-runner. In short, opponent covering consists of standing at the optimal position between the passer and receiver, and when the ball is passed going to the closest intercept point. This optimal position is learned with the batch-runner. Opponent covering performs well in the generated test cases. Some specific algorithms counter attacking strategies which have not been implemented yet, therefore it was not possible to evaluate those algorithms. They are theoretically optimal however.

Robot soccer competitions have been around for a while and have been a great environment to develop AI algorithms in. One of these environments is the AI world cup. The AI world cup environment is a virtual environment where two teams with five robots each play a soccer match. This paper focuses on defending the attacker that is carrying the ball in the AI world cup environment. This is achieved by comparing two approaches, a rule-based algorithm, and a Deep Deterministic Policy Gradient rein- forcement learning algorithm. The two algorithms are based on a coaching framework build by the TU Delft. This framework implements the actions of the robots, such as moving, jumping, and kicking the ball, and the communication between the robots. The two approaches have been evaluated and the rule-based outperforms the reinforce- ment learning algorithm. Furthermore, a teamwork strategy between the goalkeeper and defender has been developed, where this teamwork is a significant improvement over the non-teamwork algorithm.

This work analyses context effect in the evaluation of music similarity performed by human annotators to better understand the impact of context effects in the current annotation protocol of Music Information Retrieval Evaluation eXchange (MIREX). Human annotators are known to be subjective when giving similarity judgements. The Audio Music Similarity task in MIREX uses human annotators to collect similarity judgements. The annotator gives judgements to a list of candidate songs that are similar according to the participating system. The annotation protocol has no clear guidelines, and on top of that, literature shows psychological effects which can influence the similarity score. Studies show that disagreement exists between different annotators in the Audio Music Similarity task. It is argued that the disagreement is due to the natural subjectivity of human annotators, but how much of the subjectivity is natural?

In this work, context effects are explored, which are the over- or underrating of candidate songs due to specific properties of the annotated list of candidates. The properties of the list of candidates are called factors and will be used as dependent variables. The exploration of context effects is split into two parts, 1) recognizing context effects and 2) measuring the impact of the context effect. New similarity judgements are collected through crowdsourcing, this data is checked on reliability before analysing the context effects. For recognizing context effects, the changes of previous judgements made by annotators are taken as a metric to see if the annotators are noticing potential context effects. The second part is measuring the magnitude of the over- or underrating by looking at the distance of the set of judgements to the ground truth. Hypotheses are made for the dependant variables change and distance, based on the factors Order, Trend, Location, Spread and Outlier. It seems that the collected data shows signs of context effects, with the Trend and Outlier hypotheses being in line with the data. The Order hypothesis seems to be the opposite of the data. When changes are made by an annotator, the final scores of the judgements are closer to the ground truth than before the changes. However, throughout the work, no significant results are found related to context effects.

In our increasingly digital society, we are making a growing amount of data available to computers, networks and third parties. As a consequence, our sensitive data is in danger of getting exposed. The field of multi-party computation attempts to mitigate this by studying protocols that enable parties to perform their operations digitally, without the risk of privacy-violating data leaks. Among those operations are multi-party private set and multiset operations. In such a scenario, multiple parties, each with their own input set or multiset, want to collectively find the result of an operation over their inputs, without revealing these original inputs. Such operations are the cornerstone of many complex privacy-preserving protocols. For example, a two-party private set intersection forms the key to several privacy-preserving contact tracing protocols.

While multi-party private set and multiset operations have been studied for almost two decades, these privacy-preserving alternatives are often impractical: one limitation is that, to the best of our knowledge, all known protocols require several interactions between the cooperating parties. This means that rather than simply submitting their input, each party must actively take part in the protocol. In this thesis, we propose the first non-interactive protocols for privately computing set and multiset operations between multiple parties,
which rely on two constructions for non-interactive secret sharing. In addition, for operations that cannot be trivially performed using our non-interactive primitives, we propose minimally-interactive alternatives that instead rely on a homomorphic cryptosystem over elliptic curves. By using elliptic curves, this cryptosystem is faster and requires less bandwidth than the commonly used cryptosystems over integers, while retaining the same level of security. We provide proof-of-concept implementations of exact and more efficient approximate protocols that take on the order of seconds to minutes to compute, depending on the number of parties and possible inputs. Finally, we give formal proofs for the security of these protocols, so as to offer practical and provably privacy-preserving alternatives to otherwise sensitive operations.

Due to the rapid growth of diseases like diabetes, machine-learning (ML) advanced clinical decision support systems (CDSS) that support doctors or patients in their care might prove to be a valuable asset. In the battle to make healthcare more accessible, adopting ML techniques could be supportive in increasing the overall
effectiveness and efficiency of healthcare. However, the development of a CDSS is troubled by multiple problems.

Online advertisement is a multi-billion dollar industry that constitutes a primary source of income for most publishers offering free content on the Web. Online behavioural advertisement refers to the practice of serving targeted ads to online users based on their potential interests. In order to infer these interests, online advertisers aggregate and process vast amounts of behavioural data collected from the browsing activities of Internet users. Until recently, online advertisers have been using a device-centric approach to studying user behaviour and delivering targeted ads. With the development and widespread adoption of smart devices, it has become commonplace for a person to own and browse the Web from multiple devices, and online advertisers have been quick to adapt.

Cross-device tracking (CDT) constitutes the practice of identifying and tracking the user of a device rather than the device itself. This is usually achieved by grouping devices based on the likelihood that they belong to the same user. By employing CDT, advertisers are able to build more comprehensive user profiles based on a user's overall online behaviour, and serve advertisement tailored to a user's interests on all of their devices.

While these online advertising practices greatly benefit both advertisers and their clients, the privacy of online users is compromised by the amount and nature of the data collected in the process. Various solutions have been proposed over the years for performing online behavioural advertisement in a privacy-preserving manner, while no such privacy-conscious technological alternatives have been designed to address CDT practices, even though the data collected for the purpose of CDT has considerable overlap with the data collected for serving targeted ads.

In this thesis, we aim to offer a technological solution to the privacy-issues posed by the collection and processing of customer data for the purpose of cross-device tracking. To this end we design PCDT, a protocol that uses fully-homomorphic encryption and keyed hash functions to implement both deterministic and probabilistic CDT techniques that operate on encrypted data.

PCDT operates in a two server setting, where both servers engage in privacy-preserving computations to perform CDT, while simultaneously concealing the secret device data from each other through cryptographic means. The security of PCDT is based on the semi-honest security model, where parties attempt to learn as much as possible from the information presented to them, but do not deviate from the protocol. PCDT performs deterministic CDT in a privacy-preserving manner by concealing the relevant device data using a keyed hash function. The deterministic hashes allow for the fast association of devices through deterministic CDT, while the secrecy of the hashing key ensures the secrecy of the plaintext data. To perform privacy-preserving probabilistic CDT, PCDT uses fully-homomorphic encryption to train and evaluate gradient boosting decision tree models on encrypted device data. To the best of our knowledge, PCDT is the first protocol designed to perform privacy-preserving CDT.

The value of data has increased enormously over the last couple of years. Many datasets contain valuable information that can, for example, be used to make forecasts. In this thesis, the dataset of a company in the self-storage industry is analyzed. The company offers customers rental storage facilities, such as lockers, rooms, containers at several locations in the Netherlands. The rooms, also called units, are characterized by three features: location, volume, and floor. The dataset contains for each storage facility the entire renting history of each self-storage unit. For the company, it is of interest to understand and predict the demand of its customers, such that it can flexibly adjust the prices of the rooms and services based on demand. Beyond, forecasting users' demand for various types of products is a common and essential problem in many different domains (e.g., recommendation systems, transportation systems).

This thesis aims to predict the demand for the next week by applying white-box and black-box models. The problem is represented as a temporal weighted bipartite network prediction problem. Specifically, the goal is to predict the network structure at a time T+1 based on the bipartite network observed at time T-k+1, T-k+2, ..., T, where k is an integer and needs to be optimized such that the prediction error is minimized. By analyzing the data in its temporal dimension, using the autocorrelation and cross-correlation among different storage locations, floors and volumes, it was shown that the autocorrelation is high and the cross-correlation is low. This suggests that the temporal bipartite network is possibly predictable.

We have explored different state-of-the-art predictive techniques. Markov chain model, LSTM, and ConvLSTM have been selected because of their fundamental difference in the way they learn and predict. A performance comparison is given where the techniques have been applied on the storage data, and it shows that LSTM outperforms the Markov chain and ConvLSTM based on the following evaluation metrics: RMSE, MAE, and accuracy. According to our dataset, higher predictability was achieved when only the data of a single link was exploited. The Markov chain and the LSTM utilize the information of a single link to predict. On the contrary, the ConvLSTM utilizes the information of the entire network to predict. The low cross-correlation between the links explains why the LSTM outperforms the ConvLSTM. The ConvLSTM tries to capture spatio-temporal dependencies, while this, in general, does not contain much valuable predictive information. Thus, the model is introduced to more noise, making it harder to predict accurately. The LSTM also outperforms the Markov chain model, which is used as a baseline method. This proves that it is beneficial to use a complex deep learning model for this dataset to predict. However, the Markov chain performs comparable to the ConvLSTM, showing that a black-box model does not always outperform a white-box model. This emphasizes that the most suitable predictive algorithm depends on the statistical properties of the dataset.

The theoretical upper bound of the predictability of the network is computed. It is the upper bound that can be used to compare the realized performance to the maximum achievable prediction performance for any predictive algorithm. The difference between the performance of the best performing algorithm to our dataset, LSTM, and the theoretical upper bound is still large, indicating that there is still room for improvement.

Privacy-preserving data aggregation protocols have been researched widely, but usually cannot guarantee correctness of the aggregate if users are malicious. These protocols can be extended with zero-knowledge proofs and commitments to work in the malicious model, but this incurs a significant computational cost on the end users, making adoption of such protocols less likely.

We propose a privacy-preserving data aggregation protocol for calculating the sum of user inputs. Our protocol gives the aggregator confidence that all inputs are within a desired range. Instead of zero-knowledge proofs, our protocol relies on an asynchronous probabilistic hypergraph-based detection algorithm with which the aggregator can quickly pinpoint malicious users. Our protocol is robust to user dropouts and is non-interactive apart from the registration phase. We describe several optional extensions to our protocol for temporal aggregation, dynamic user joins and leaves, and differential privacy. We analyse our protocol in terms of security, privacy, and detection rate. Finally, we compare the runtime complexity of our protocol with a selection of related protocols.

Elliptic Curve Cryptography (ECC) performance is a major performance bottleneck when serving many VPN clients from a single server on a low-frequency FPGA softcore CPU. Using an area-efficient Elliptic Curve Point (ECP) multiplication accelerator core on the same FGPA, a much higher amount of clients can be served using the same FPGA chip. Using the accelerator core, the obtained speedup ranges from 1.6x in a suboptimal configuration up to 7x with a configuration that maximizes the use of ECC when connecting new clients to the server. In this optimal configuration, the total amount of clients that can be served by a single OpenVPN server increases from 80 in the base case, to 350 in the accelerated case.

For a system to be able to interpret data, learn from it, and use those learnings to reach goals and perform tasks is what it means to be intelligent [1]. Since systems are not a product of nature, but rather made by humans they are called Artificial Intelligence (AI). The field of Side-Channel Attacks (SCA) has benefited from applying AI systems to their problems. Operations previously to resource-intensive to perform can now be executed using AI. Currently, the focus lies on exploring which parameters result in the optimum performance when classifying side-channel traces. But since this application has only recently been applied, there is much more research to be done. As of now, the literature claims that a reduction in the size of the architecture would result in an improvement of the explainability of the models used. However, this change in explainability has not been explicitly proven to hold for SCA models. This created a gap in knowledge. This paper aims to close this gap by exploring these assumptions. The goal is to explore if a reduction in complexity of SCA models leads to improved explainability. An experiment was conducted using two existing SCA architectures with a small and large complexity respectively. Using heatmaps, the explainability of these models were assessed to investigate the existence of patterns. The results show a difference in the consistency of the classification process, where the model with the lowest complexity could more consistently state why a certain classification was made. The results indicate that the explainability of a given SCA model can be improved by decreasing its complexity.

The Internet consists of many subnetworks, which are connected to each other. These subnetworks are the autonomous systems (ASes) that make up the Internet: each hosts a part of it. In order to successfully determine routes from one of these ASes to the other, the Border Gateway Protocol (BGP) is used. This protocol has several security flaws however, and exploitation of them has lead to parts of the Internet being temporarily unreachable. In order to combat these flaws, several security solutions have been developed already. However, none of these have been deployed on a wide scale yet. As such, this thesis focuses on the question: why not, and what can be done to protect BGP in the future? This thesis includes an analysis of the BGP threat landscape, to find which threats are most relevant, and to find out whether or not solutions have adapted to the threat landscape. It also includes a comparison of solutions on different practical security aspects. From this comparison, I found that no solution is able to prevent attacks if only one autonomous system deploys it. Due to this, I suggest to shift attention to detective security. This thesis also includes an analysis of some detective security schemes, to see which properties of these schemes can be used for a new scheme. This new scheme is designed to comply with a list of requirements, and it uses properties from three other schemes. Development of this scheme is left as future work. Altogether, this thesis should provide a new direction for the future of BGP security.

Inter-Autonomous System (AS) route monitoring is the process of collecting the inter-AS routing information. This information flows on the Internet in the form of BGP UPDATE messages, and the BGP data are the messages obtained by the monitors. Existing methods of monitor placement rely on the network topology which provides inadequate visibility. This thesis proposes a novel scheme for monitoring the BGP data by selecting monitoring locations based on BGP message flow.

In the past years, society has become increasingly more reliant on the Internet. Consequently, the security of the Internet became of critical importance. This thesis focusses on the security of one of the Internet's main protocols. This protocol, called the Border Gateway Protocol (BGP), is used to exchange information that allows Internet traffic to reach its intended destination. BGP is vulnerable to misconfigurations and attacks that can cause a range of problems. This thesis focusses on one of them: BGP origin hijacks. In this thesis, a year of possible origin hijacks is analysed. These possible origin hijacks were detected by BGPStream between 20 May 2018 and 31 May 2019. Analysing these hijacks gives insight into the causes and characteristics of origin hijacks. This can help to find the most pressing issues and may provide guidance in securing BGP. Various data sources are used to collect and compute features that give more information on each hijack. These features are used to find relations between hijacks and to label them using labels that indicate a cause or a certain aspect of the hijack. These relations and labels are used to analyse groups of similar hijacks. This approach is very effective. Using the context of a group of hijacks gives much more insight than looking at hijacks individually. It shows that many of the possible hijacks are likely not a hijack at all and that hijacks that look like origin hijacks are often the result of another type of attack called a path hijack. In addition, this thesis provides a way to detect several types of misconfigurations and points out weaknesses in the detection system used by BGPStream. It also gives an overview of the characteristics of hijacks and how often specific behaviour occurs.