We increasingly encounter artificial intelligence-based technology in our daily lives, from smart home devices to self-driving cars to invisible systems running on our internet. Many artificial intelligence techniques use machine learning, algorithms that learn to predict or act based on collected data. Unfortunately, the most popular machine learning techniques, such as neural networks and ensembles, are so complex that humans cannot understand how they make predictions. Without understanding the prediction process, it is difficult to trust the model. Therefore, in this dissertation, we work on algorithms that learn models that are understandable to humans. The type of model we consider is a decision tree, a flowchart-like model that can easily be visualized so humans can understand it. These models ask a series of questions about an input and use the answers to derive a prediction.

Decision trees were popularized in the 1980s and extensively studied, but there is still room for improvement. The most popular algorithms for learning decision trees are fast but do not necessarily lead to the best performance. They are not robust, meaning tiny changes in the data can negatively influence the quality of their predictions. Also, the existing algorithms cannot be directly applied to problems where multiple sequential predictions have to be made. Therefore, this dissertation studies several techniques for learning decision trees for robustness and sequential decision making problems.

In Part I of the dissertation, we consider the problem of optimizing decision trees to make good predictions while being robust to small changes in the data. In Chapter 4, we tackle the problem of learning good decision trees quickly in this setting. We improved the runtime of an existing algorithm by speeding up one of the key operations. In Chapter 5, we solve the problem of finding the best possible robust decision tree. The idea is to formulate the problem as an Integer-Linear Program, a special mathematical problem that can be solved with highly optimized algorithms. In Chapter 6, we propose a method that allows learning of models that are more flexible in terms of robustness, i.e., by allowing data changes in different shapes. To create an efficient algorithm, we optimize only the model’s predictions, not the model’s question part. Finally, in Chapter 7, we use techniques for improving data privacy to enable robustness against another kind of data change: someone adding or removing data.

Part II of this dissertation is about sequential decision making problems. In these settings, we control a device or agent that tries to achieve some goal in a potentially uncertain environment. Sequential decision making problems are significantly different from the supervised learning problems considered in Part I since the data is not pre-collected. This class of problems encompasses many real-life problems; one of the simplest of those could be a thermostat that measures the temperature in a room and needs to decide whether to turn a heater on or off constantly. Highly complex problems such as self-driving cars can be modeled similarly. We aim to find a controller represented by a simple decision tree for such problems. Such a controller is called a policy, and by representing it with a small decision tree, humans can understand it. In Chapter 9, we assume that we have a perfect mathematical description of the problem and use it to find the best possible decision tree via Integer Programming techniques. Later in Chapter 10, we assume that we can only interact with the environment and do not have a mathematical description of the problem. In this setting, we find good policies by iteratively updating the tree to achieve better scores using gradient information.

In our research, we have developed various algorithms for learning decision trees in settings that are hard to optimize with existing methods: robust predictions and sequential decision making. We hope that our work on decision tree learning for these settings allows human-understandable machine learning to be used in more real applications in the future. By improving model understanding and robustness, we aim to enable machine learning systems that humans can trust.
@en

In modern society, data is increasingly important for people's daily lives and commercial activities to benefit organization, management and analysis. Proper data utilization involves different stages, and the data life cycle is introduced to describe the procedures as generation, collection, processing, storage, management, analysis, visualization and interpretation. Along the data life cycle, data privacy is highly concerned since possible breaches can lead to the abuse of personal information and financial loss. In this thesis, we advance data privacy protection in data processing, data management, and data analysis correlated with the Spark Living Lab project in the domain of supply chains and machine learning. To enhance privacy protection, we propose solutions based on differential privacy and cryptographic protocols since they provide strong and provable security and privacy guarantees. Moreover, we integrate differential privacy and cryptographic protocols to achieve strong privacy guarantees efficiently and practically for data processing, management, and analysis.

In data processing, we focus on data anonymization and location data perturbation in supply chains. Data de-identification is essential to comply with privacy laws, such as GDPR, before possible sharing or analysis. We propose an anonymization algorithm by combining differential privacy and k-anonymity to achieve stronger privacy guarantees or better data utility than using them alone. Meanwhile, we consider trajectory hiding under possible attacks and real maps, which propose a more practical solution to share trajectory data under privacy protection.

In data management, we address secure data sharing with cryptographic protocols. Data sharing is vital in data management to advance collaboration and knowledge. However, possible data breaches and malicious inputs can lead to potential financial loss and identity theft. In this thesis, we propose a framework for sharing logistic data in a privacy-preserving way using blockchain and cryptographic protocols. Differential privacy is applied to anonymize data, while cryptographic protocols enhance privacy during data transmission.

In data analysis, we pay attention to privacy-preserving machine learning. Machine learning models are usually trained on large datasets which may contain sensitive personal information. It is important to consider privacy protection during the training and utilization of models. We use differential privacy and secure multi-party computation techniques to design a framework for collaborative learning among multiple parties against inference attacks. Also, we utilize zero-knowledge proof to validate model integrity without leaking the model.@en

In recent years, the practice of risk-assessment has started to utilize machine learning to take in larger data sets and improve prediction accuracy. Simultaneously, it has expanded into the sentencing stage of the criminal justice system. This paper analyzes the effect of these developments on the consideration of the theories of punishment during sentencing. After first going over key characteristics of both the theories of punishment and the practice of risk-assessment, it presents a series of connections between aspects of machine learning enabled risk-assessment and each of the theories of punishment. It then provides developers of such systems with both general advice and advice aimed specifically at the highlighted effects, rooted in various design methodologies such as Systemic Design and Value-Sensitive Design, in order to better account for these effects. Future research can expand upon this paper both in depth and in scope: in depth by carrying out the empirical research needed to verify and improve upon the ideas in this paper; in scope by extending this research to other ethical debates surrounding machine learning in criminal justice.

The internet traffic is constantly rising nowadays due to the significant increase of the devices connected to the Internet. As a consequence, many cyber risks have arisen. Cybercriminals are trying to exploit the vulnerabilities of these devices to cause damage and gain profit. Monitoring the network traffic and detecting such threats has become essential in order to keep safe systems that are connected to the Internet. The powerful properties of state machines and the sequential nature of the network traffic data, makes them an interesting and promising solution for the implementation of an intrusion detection system.

The goal of this thesis is to implement a new state-merging heuristic which will speedup the state machine building procedure without a significant loss on the quality of the model, and use it to detect malicious host on network traffic data. The new state-merging heuristic is utilizing the Locality-sensitive Hashing concept to store the future traces of each state and simplify the consistency check for the merge of two states. The network traffic data used are in the NetFlow format, and they are encoded and converted into traces in order to build the state machine model and measure its performance. The state machine built is modeling a malicious behavior and used to classify other hosts.

We show that the models built can effectively detect the malicious hosts, with its performance being comparable to the one of a state-of-the-art model. At the same time, the time needed to build the model is much less when compared to the time needed by other state-merging heuristics.

The Internet is a relatively new technology that the world has become immensely dependent on. It is a tool that makes it possible to simplify our lives and better our society. But as with many things, there are people who with to exploit this tool we have for their own malicious gain. One of the mechanisms that we can use for protection against these malicious actors is the intrusion detection system. Machine learning-based intrusion detection systems (IDS) have been heavily researched for a number of years now. Much of this research, though, appears to be conducted using improper methodologies and incorrect evaluation.
Such methodologies include training and testing IDSs with unrealistic data and using uninformative metrics to determine performance. In this research, we perform a case study using one such IDS. This IDS is trained and evaluated using real network traffic collected from a real-world network. Additionally, we test its performance on actual attack traffic. This research demonstrates that an IDS that is trained with unrealistic data performs nowhere near as well as is claimed by the author when trained using real network traffic. Finally, we propose CAML-IDS, a framework for the correct assessment of machine learning-based intrusion detection systems. This framework can assist future IDS research by preventing incorrect evaluation, in turn preventing the formulation of incorrect research.

Strape enables users to initiate and authorize peer-to-peer transactions without alphanumerical input by
using gestures. OK, a smart wallet app, combines authentication, marketing and payment services in a single
app. By choice, peer-to-peer payments were excluded from the initial design of the OK app, because no
compelling user experience for this type of payment was available, yet. With Strape there is. To incorporate
the Strape functionality inOKand other host applications, the Strape functionality needs to be made available
as an SDK.
The goal of the project is therefore twofold: An SDK must be developed that provides the gesture-based
functionality and adheres to a number of integration use cases for various possible hosting applications. First,
it should be possible to initiate transactions by using gestures, and authorize the transactions by tapping on
a discovered user and swiping it towards them. Secondly, the SDK should be structured in such a way that it
is possible for a host application to fully integrate the SDK.
The system is implemented using a set of configurable classes with the business logic and a data provider
that handles all data mutation and persistence. Furthermore, it holds views that can be extended by a host
application for further customization. Discovery is implemented by using Bluetooth Low Energy, a protocol
that enables constant advertisement and scanning of users in proximity. To ensure testability and maintainability,
aModel-View-Presenter architecture is used.
The SDK is implemented in OK at various points in the app, namely as a favorite between payment accounts,
IDs and tickets, as a shortcut in the main overview and as an overlay with transaction objects. OK
mostly uses the standard SDK configuration variables.
The project was managed using Scrum with 1-week sprints. Version control was handled by Git and
SemVer and the project was continuously integrated by using Bitrise. Although a connection to the Strape
Virtual Account System was planned, time constraints and prioritization of the core Strape functionality resulted
in leaving it out of the final product.
The result satisfies the project goal in that an SDK is realized comprising the core Strape functionality
with the specified integration requirements. Furthermore, a first version of the SDK is integrated in the OK
app, satisfying all UI integration requirements. The project has showed us the importance of the necessary
overhead with software development projects such as this one. We recommend developing the requirements
we deemed out of scope for this project to complete the entire peer-to-peer transaction proposition.

The "telephone'' model has been, for the last one hundred thirty years, the base of modern telecommunications with virtually no changes in its fundamental concept. The arise of smaller and more powerful computing devices have opened new possibilities. For example, to build systems able to give to the user the illusion of being talking to the remote party as if both where in the same place. To achieve this still many challenges have to be overcome. In this thesis, a part of the acoustical signal processing problem is treated. To acoustically create the illusion of presence, fast and accurate control over the sound field in a room is required. The sound field given one or more sources is subject to different acoustical phenomena, such as reflection and diffraction. Because of these, to model or estimate the sound field in a room is in general a difficult task. In particular acoustical reflection poses an important challenge. The sound field reflects on the walls, ceiling and floor and a moment later those reflections reflect again, and later these reflect again. This recursive process makes the number of reflections as a function of time to increase, in general, at a geometric rate. To synthesize an artificial sound field in real time, one has to be able to model these reflections fast and accurately enough. In this thesis a fast algorithm to model the sound field in box-shaped rooms is proposed. Part one of this thesis begins with an introduction to the topic, here the different acoustical phenomena of interest are explained, and the concept of room impulse response (RIR) is introduced. The RIR is defined as the time-domain signal sensed at a receiver position as generated by a point source that emits an impulse. Assuming a linear time-invariant (LTI) model, if the point source emits not an impulse but an arbitrary signal, the actual sound field at a given observation location can then be modeled as a convolution of the source signal with the RIR. Moreover, since we are assuming a linear model, the sound field generated by an arbitrary number of point sources emitting arbitrary signals can be easily computed once the RIRs from the locations of the sources to the observation locations are known. Efficient computation of the RIR is therefore of theoretical and practical interest. Consequently, this part concludes with a summary of the most prominent algorithms to simulate the RIR. Part two of this thesis contains the relevant papers that make up this work. The analysis is given first for the case of fully reflective walls. It is noted that in this case all the acoustical reflections can be modeled by a set of virtual sources following a periodic structure over a lattice. The whole set of virtual sources is generated by the repetitions of a small set of sources called "the mother sources''. On the other side, the Poisson summation formula establishes the relation between periodicity and discretization under the Fourier transform. Relating these concepts, it is shown that by carefully discretizing the spectral representation of the RIR of the mother sources in free-field, the exact periodic structure that makes up the sound field in a room can be obtained. This is the key idea behind the proposed method. Carefully discretizing all domains, and making use of the fast Fourier transform (FFT), a fast multichannel RIR simulation method is obtained. Unfortunately this idea only works for fully reflective walls. By allowing the walls to have constant complex-valued reflection coefficients (this is, to model absorption and phase shift at the walls) the sound field of the set of virtual sources is not anymore periodic. A generalization of the Fourier transform is then introduced. First, a generalized Poisson summation formula is derived. This formula relates discretization in the generalized Fourier domain to a geometrically weighted periodic summation in the reciprocal domain. Basic properties of this transform are derived, its application to non zero-padded linear convolution is derived, but moreover a fast implementation, called the generalized fast Fourier transform (GFFT), is given. The proposed method is then extended to account for walls with constant complex-valued reflection coefficients. It is shown that by separating the sound field of the mother sources into its orthant-sided parts (the analogue of the single-sided parts of a function of a scalar variable), the sound field inside a room can be expressed as a sum of geometrically weighted sound fields generated by the periodic set of virtual sources. This summation is then related to a sampling condition on the generalized spectrum of the orthant-sided parts of the sound field of the mother sources. Using the GFFT the method simulates the RIR given a source at a dense set of spatial positions with very low complexity. In the experiments a comparison with a model called the mirror image source method (MISM) is given. In one scenario, the time the MISM would take to compute the RIR at a dense set of positions is estimated to be about one and a half years. The newly proposed method computes the RIR at all positions in only forty-eight minutes. This shows the contrasting difference in computational complexity, making the new method an important step on the road to simulate realistic sound fields in real time.@en