Search

Polar Mesospheric Summer Echoes (PMSE) are strong coherent radar echoes that occur in the 80 to 90 km altitude range of the mesosphere during the Arctic summer months. These echoes are of significant interest to the space physics community as they provide insight into changes that occur in the atmosphere. To better understand these changes, large datasets of PMSE echoes need to be analysed. In this study, we aimed to develop a deep learning model that could segment PMSE signal data for analysis on larger EISCAT VHF datasets. For the task, we employed a UNet and a UNet++ architecture and tested how pretrained weights from other source domains perform. Next, different loss functions were tested and last the novel object-level augmentation method ObjectAug was employed with other image-level augmentation methods to increase model performance and reduce potential overfitting due to a small training dataset. The results indicate that using randomly initialized weights was the better option for the PMSE target domain and that the use of different loss functions only had a small impact on model performance. When using image- and object-level augmentation the best performing model was reached. It was also seen that there exist inconsistencies in the PMSE signal groundtruth labels. Dividing the inconsistencies into two categories: Granular and Coarse, it was seen that using object-level augmentation had a significantly higher performance on the Granular labelled PMSE signal samples. Overall, our study indicates that the best performing model can be used to segment PMSE for larger datasets or as a supportive tool for further labelling of PMSE signal data.

This thesis explores integrating deep learning techniques into photography, aiming to automate the identification of good images within large datasets. The primary focus is developing a deep learning-based system called DeepRoom that rates and evaluates photographs based on photography-specific technical criteria. To accomplish this, the research methodology encompasses qualitative research alongside developing a system prototype. A section overviews deep learning, photography, and related work and emphasizes its relevance to the research objectives. Implementation details include describing development tools and processes employed to construct the deep learning models and curate the dataset. These models' performance is assessed in the following evaluation phase, and a comparative analysis is conducted against existing software solutions. Encouraging results are observed, particularly in object detection and exposure classification, while identifying areas for improvement, such as refining the blurry and skewed horizon models. In conclusion, this research highlights the contributions of DeepRoom and proposes future work, including dataset expansion and model refinement, to enhance its capabilities further.

Presentation at the Dataverse Community Meeting 2023, 05.06. - 07.06.23, University of Minho, Braga, Portugal

In the past decade, substantial advancements have been achieved in effectively utilizing video surveillance and associated analysis technologies within the realm of sports. This progress has been particularly noteworthy in elite sports, where the exploitation of athletes’ digital footprints for sports analytics has emerged as a catalytic factor, ushering in a paradigm shift in comprehending and formulating strategic approaches to the game. The architecture of sports video analytics systems can be broadly categorized into (1) tagging and (2) analysis. Tagging involves annotating metadata to specific video sequences and events, and this tagged metadata is subsequently utilized in the causal analysis process. Multiple enterprise solutions are available today for recording videos, and positions and producing tagged data for the top teams. The issue is that they are often expensive, time-delayed metadata, and the sports organizations do not control where the data is stored or how the analytics company uses it. The alternative to enterprise solutions is manually generating the soccer metadata, which is time-consuming and possibly impossible if, for example, one wants to tag every player’s position throughout a game. This thesis presents Mearka, a distributed soccer tagging system based on cheap common-off-the-shelf components. It allows for tagging events LIVE during a soccer game through the Mearka-app, as well as generating player position metadata with time offsets into a user-uploaded video through the Mearka web-interface, automatically detected using machine learning. After detection, it is possible to download the soccer metadata as a JSON file through the web-interface. The experiment results demonstrate that Mearka can complete the detection of players’ positions from a 90 minutes soccer game within 12 hours after detection is started, with a video resolution of 1920x1080 at 25FPS. Expanding Mearka to only detect on every 10th frame could potentially make Mearka a viable real-time tagging option, as it is able to detect on ≈3 frames per second, and a turnaround of 12 hours detects on every single video frame.

Presentation at the Dataverse Community Meeting 2023, 05.06.23 - 07.06.23, University of Minho, Brage

IoT is a field of technology of ever growing importance in our daily lives. From smart cities, health devices, climate observations, appliances, and so much more, IoT surrounds us now more than ever. The types of devices being added to IoT networks is ever growing, and as this variety of hardware and software increases, so does the difficulty of working with them. Ensuring inter-compatibility between devices, testing new communication protocols, and writing software for emerging technologies becomes a complex challenge. To help solve this challenge are IoT Testbeds. IoT Testbeds help developers, researchers, and many more groups of people explore and test their IoT solutions in contexts of real IoT Devices. These testbeds exist today, but as far as we know, no Jack of all trades testbed exists that supports all features one might want from a testbed. This thesis will introduce a first draft of a new testbed. Introducing a system design, architecture, and implementation that theoretically and practically implements all these features. Also highlighting issues with this design and ways to tackle them. In the end contributing a foundation onto which a powerful system could be built. The challenge the thesis aims to tackle is, in short: What are the needed features that make up a good testbed? And how can we incorporate these features into a simple, flexible, unified system?

In the context of computer networks, decentralization is a network architecture that distributes both workload and control of a system among a set of coequal participants. Applications based on such networks enhance trust involved in communication by eliminating the external author- ities with self-interests, including governments and tech companies. The decentralized model delegates the ownership of data to individual users and thus mitigates undesirable behaviours such as harvesting personal information by external organizations. Consequently, decentral- ization has been adopted as the key feature in the next generation of the Internet model which is known as Web 3.0. DIDComm is a set of abstract protocols which enables secure messaging with decentralization and thus serves for the realization of Web 3.0 networks. It standardizes and transforms existing network applications to enforce secure, trustful and decentralized com- munication. Prior work on DIDComm has only been restricted to pair-wise communication and hence it necessitates a feasible strategy for adapting the Web 3.0 concepts in group-oriented networks. Inspired by the demand for a group communication model in Web 3.0, this study presents Zero- Comm which preserves decentralization, security and trust throughout the fundamental opera- tions of a group such as messaging and membership management. ZeroComm is built atop the publisher-subscriber pattern which serves as a messaging architecture for enabling communi- cation among multiple members based on the subjects of their interests. This is realized in our implementation through ZeroMQ, a low-level network library that facilitates the construction of advanced and distributed messaging patterns. The proposed solution leverages DIDComm protocols to deliver safe communication among group members at the expense of performance and efficiency. ZeroComm offers two different modes of group communication based on the organization of relationships among members with a compromise between performance and security. Our quantitative analysis shows that the proposed model performs efficiently for the messaging operation whereas joining a group is a relatively exhaustive procedure due to the es- tablishment of secure and decentralized relationships among members. ZeroComm primarily serves as a low-level messaging framework but can be extended with advanced features such as message ordering, crash recovery of members and secure routing of messages.

One of the novel new features of mobile 5G networks is what is commonly known as "Ultra Reliable Low Latency" communication. To achieve the "Low Latency" part, it is necessary to introduce processing and storage capabilities closer to the radio access network, thus introducing Edge data centers. An Edge data center will be capable of hosting third-party applications and a user of these applications can access them using the cellular mobile network. This makes the network path between the user equipment (UE) and the application short in terms of physical distance and network hops, thus reducing the latency dramatically. This thesis looks into these new features of the 5th-generation mobile networks to establish if, and how they can be used to orchestrate container-based applications deployed at edge data centers. The orchestration mechanism suggested will be described in more detail in the thesis body but as an overview, it involves using the user's positions and the knowledge about which applications the users are accessing and information about where these applications reside to move applications between edge data centers. One of the 5G exploration findings was that the location of users in a 5G network can be determined using the Network Exposure Function (NEF) API. The NEF is one of the new 5G network functions and enables trusted third-party actors to interact with the 5G core through a publisher-subscriber-oriented API. The proposed orchestration strategy involves calculating the ``weighted average location'' of 5G users who have accessed the specific application residing in the Edge within a specified time frame. A live 5G network with a stand-alone (SA) core was not available at the time of writing and part of the thesis work has therefore been to identify if there exist network emulators with the functionality needed to reach the goal of this thesis, i.e. design and implement the orchestrator based on interaction with the network. More specifically: can we find a NEF emulator that can be configured to give us network data related to user equipment location? Unfortunately, the three alternatives considered: Open5Gs, NEF\_emulator, and Nokia's Open5Glab do not fully meet our requirements for generating user events. Open5Gs an open source 5G network implementation lacks the whole NEF north-bridge implementation, NEF\_emulator has limited implementation and integration complexities, and Nokia's Open5Glab's simulated users are inactive and thus do not generate sufficient data. Given the absence of suitable emulators to generate the needed data, the thesis pivoted to also include the design and implementation of a mobile network emulator with the following key components: a mobile network abstraction that encompasses crucial elements from 5G, such as users and radio access nodes, allowing users to connect to the mobile network; a network abstraction that hosts emulated edge data centers and the corresponding applications accessible to connected users; and mobile network exposure that exposes mobile network core events through a simplified NEF north-bound API implementation. Finally, the thesis concludes by implementing the proposed orchestration strategy using the mobile network emulator, demonstrating that orchestrating can effectively reduce the end-to-end latency from users to applications, as evidenced by the obtained results.

The Norwegian fishing industry faces a significant issue of fishery crimes, with product traceability systems presenting a potential solution to counter these illegal activities. Current supply chain management in the seafood industry is vulnerable to information alterations, thereby complicating product traceability. Blockchain technology, with its unique properties, offers an interesting approach to address these challenges. Despite this, existing blockchain-based product traceability systems often fail to integrate government regulation and provide limited access to traceability data for consumers. Moreover, those providing such access often lack user-friendliness. This thesis explores if a blockchain-based product traceability system can support supply chain management, enhance consumer confidence, and enforce regulatory compliance. We conducted a review of existing literature and assessed the potential of blockchain technology to optimize supply chain management. Furthermore, a traceability system, entitled SeaChain, incorporating a permissioned blockchain, smart contracts, and governmental regulations was developed. We evaluated this system and compared it with existing systems. Our findings suggest that blockchain technology can enhance supply chain management, bolster consumer trust, and aid in mitigating fishery crimes. The research conducted provides valuable insights for improving supply chain management and contributes to future studies in this field.

Annually, avalanches claim an average of 100 lives[9]. In Norway, seven people die and 67 get injured on average each year due to avalanche accidents[15]. For those who survive the trauma but are buried under snow, time is of the essence. This master’s thesis aims to provide an efficient solution for avalanche victim search during rescue operations. Traditional methods rely heavily on human intervention, which can result in delays and loss of critical information. To overcome these challenges, unmanned vehicles, like drones, have emerged as a promising tool for initial search operations in high-risk avalanche areas. However, the successful implementation of unmanned vehicles requires a spe- cialized avalanche beacon capable of remote control, data storage, and trans- mission. The primary objectives of this thesis are threefold. Firstly, to develop a proto- type proof-of-concept of a novel avalanche beacon to demonstrate its feasibility. Secondly, to examine the impact of radio interference on the beacon’s per- formance to ensure accurate and reliable operation. Finally, to evaluate the practicality of the beacon for avalanche rescue purposes. The prototype beacon successfully meets most of the outlined requirements, demonstrating its potential as a valuable tool in avalanche search and rescue. It can detect and store digitized versions of analog radio signals transmitted by common avalanche beacons and is compact and lightweight enough for integration with unmanned vehicles. In conclusion, this thesis presents the development and evaluation of a proto- type digital avalanche beacon receiver. The findings highlight the potential of the beacon for avalanche search and rescue operations, although further refine- ment and testing are necessary before use in real-life rescue operations.

FishAI: Sustainable Commercial Fishingis the second chal-lenge at theNordic AI Meetfollowing the successful MedAI,which had a focus on medical image segmentation and trans-parency in machine learning (ML)-based systems. FishAI fo-cuses on a new domain, namely, commercial fishing and howto make it more sustainable with the help of machine learning.A range of public available datasets is used to tackle three spe-cific tasks. The first one is to predict fishing coordinates tooptimize catching of specific fish, the second one is to createa report that can be used by experienced fishermen, and thethird task is to make a sustainable fishing plan that provides aroute for a week. The second and third task require to someextend explainable and interpretable models that can provideexplanations. A development dataset is provided and all meth-ods will be tested on a concealed test dataset and assessed byan expert jury.

Fully Homomorphic Encryption (FHE) and Trusted Execution Environ-ments (TEEs) are complementing approaches that can both secure computa-tions running remotely on a public cloud. Existing FHE schemes are, however, malleable by design and lack integrity protection, making them susceptible to integrity breaches where an adversary could modify the data and corrupt the output. This paper describes how both confidentiality and integrity of remote compu-tations can be assured by combining FHE with hardware based secure enclave technologies. We provide a software library for performing FHE within the Intel SGX TEE, written in the memory-safe programming language Rust to strengthen the internal safety of software and reduce its attack surface. We evaluate a sample application written with our library. We demonstrate that we can feasibly combine these concepts and provide stronger security guar-antees with a minimal development effort.

This master thesis reports on the research of simulating hill ascensions on winter roads. By utilizing Python and Excel a simple prototype simulation was created. This simulation shows the velocity changes during the ascension, with different variables such as: initial velocity, gear, weight distribution, environment etc. The variables used are mostly based on real life measurements, with a few exceptions where educated guesses where taken. The results of the simulation shows clearly how the different variables impact the simulation. These results reflect the expected outcome and statements from other research papers. Additionally these results also proves the hypothesis: It is not only the physical aspects of a hill ascension that needs to be taken into consideration, but also how the driver is driving. Especially when it comes to what gear is being used and the impact of a gear change during the ascension. Further validation of the accuracy cannot be taken before proper experiments and potential tuning are completed.

In this thesis, we examine the viability of training a convolutional neural network using synthetic data. The cnn is used for image recognition in an rms. We create a program that can render 3D models from STL files as images with varied backgrounds. The process of creating and training an image recognition model is also automated. Lastly, the model is used for image recognition. The report compares different methods and hyperparameters used in training a model. Transfer learning is found to be suited for synthetic datasets. Using the pre-trained feature extraction layers of the VGG-16 model, we train an image recognition model with better than 90% accuracy in the laboratory. We then demonstrate the use of this model for object detection, and suggest avenues for further development.

The World Health Organization has reported that more than 80% of the world’s adolescent population is insufficiently physically active [1]. Up to five million deaths per year could be averted if the global population were more active[1]. The low adherence to physical activity shows the need to implement services that promote physical activity. In addition, it is crucial to educate people on the benefits of being physically active and the negative consequences sedentary behavior imposes. This thesis proposes a social media chatbot with the integration of an activity tracker that aims to motivate people to increase their daily step count. The chatbot, AktiMotBot, encourages people by implementing behavior change techniques in its messages and functionality. We use popular technology, such as social media applications, to ease access. Further, the use of chatbots has grown. A chatbot gives a service that is always available to the user and is cost-effective. In addition, chatbots have familiar interfaces that ease their use. Finally, activity data is integrated into the chatbot as a motivation and personalization tool, enabling monitoring behavior change. A thorough investigation of social media applications was conducted to ensure users’ privacy and security. A usability study investigated how potential users perceived the system, and the usability of the chatbot was scored as average. The results showed that the chatbot was able to increase the motivation of half of the participants. Finally, the findings from this research are that chatbots could motivate people to increase their physical activity levels and make people more aware of their step count.

Physical inactivity has been identified as a global pandemic, physical inactivity causes multiple health outcomes in different demographic group such as coronary heart disease, type 2 diabetes, colon, and breast cancer. A physical inactive person takes less than 5000 steps a day. To try to reduce physical inactivity along individual for healthy lifestyle, this thesis provides personalized digital nudge. Nudge means to guide someone to do something that is beneficial for the long-term benefit of the person being nudged and doing so using UI (user interface) in digital environment is known as digital nudging. As people are relying more on technology for their decision making, the information collected from the integration of the devices is used to provide personalized nudges. As people have access to smartphones and wearable devices, data is collected from these devices to provide tailored nudges to achieve minimum required steps to reduce inactivity. Personalized nudge is a smart nudge which predictably influence people's behaviour. It is a type of digital nudge. This kind of nudge takes user’s information into account before nudging a user. This thesis also provides recommendations (new activities) based on person’s preference. The presented system was also tested by real users, and the feedback suggested that the presented system indeed urged them to be more active.

In large simulations, like predicting the movement of ocean particles, it is common that simulation executions are related when they share one or more inputs. When the number of simulations increases, it becomes harder for users who run the simulations to keep track of all the simulations. Also, more storage spaces are wasted if there are multiple copies of the same input files. This thesis describes a system that collects data from previous simulations, allowing users to search for the data they need to run the next simulation. Also, the system identifies the same files that were used in previous simulations, which allows users to re-use these files instead of copying the files to a new simulation folder to use them. Among the simulations that were executed in our current environment, the system identifies around 11\% of input files that are shared by the simulations. Users can refer to the same file to use it instead of copying the file to new simulation folders. The conclusion is that the system helps users who run simulations to reduce their efforts and time to find input files that are used in previous simulations when they set up for a new simulation. Also, it saves storage space on the computing cluster where the simulations run on by identifying the duplicated data.

A Cochlear Implant(CI) is an implant that replaces the functionality of the inner ear with an electronic prosthesis. The prosthesis stimulates the auditory nerves within the cochlea so that people with auditory disabilities regain hearing. While the medical process of inserting the implant is relatively straightforward, learning how to use the implant may be difficult. This thesis proposes a digital training platform that can aid with the learning process for CI users and lighten the transition into the audible realm. Providing efficient training tools available at home without assistance from third parties can increase independence for CI users, as it becomes easier to conduct audio training. The results from a questionnaire performed during the thesis with CI users as participants found that participants spent, on average, one hour traveling to and from training centers or audio therapists. Introducing home-based training tools tailored toward post-lingually deafened CI users can help free up patients' time and alleviate the learning process. Therefore, a digital training platform was developed, tailored towards the needs of post-lingually deafened CI users. CI users were involved via questionnaires and semi-structured interviews throughout the development process, and their feedback was gathered to inform the design and increase usability and effectiveness. The participants' feedback and the findings from a literature review based on related research crafted the requirements specification for the proposed digital training platform. Throughout the thesis, the main goal was to create something that could be deployed and maintained, accessible and reachable for those who may use it. Therefore, a cost analysis of different cloud computing services was carried out to find the cheapest deployment options to avoid adding a monetary barrier to entry. Linode Kubernetes Engine (LKE) was an affordable and efficient option. Along with LKE, the efficacy of deployment through single board computers was investigated. The data gathered through performance testing done during the thesis indicates that the current generation of the Raspberry Pi platform could be used to provide a cheap alternative to deployment if the number of expected users is relatively low. However, a side effect of using single-board computers is that it puts the responsibility to provide the availability and reliability requirements on the developer. As a result, you can not rely on the monetary incentives of a cloud provider. Since the availability and reliability requirements of the proposed digital training platform could not be ensured when utilizing single-board computers, the final deployment utilizes LKE and is accessible through: www.lyttetrening.no

Research institutions and funding agencies are increasingly adopting open-data science, where data is freely available or available under some data sharing policy. In addition to making publication efforts easier, open data science also promotes collaborative work using data from various sources around the world. While the research datasets are often static and immutable, the metadata of a file can be ever-changing. For researchers who frequently work with metadata, accessing the latest version may be essential. However, this is not trivial in a distributed environment where multiple people access the same file. We hypothesize that the publisher subscriber model is a useful abstraction to achieve this system. To this, we present Gurret: a distributed system for open science that uses a publisher-subscriber based substrate to propagate metadata updates to client machines. Gurret offers a transparent system infrastructure that lets users subscribe to metadata, configure update frequencies, and define custom metadata to create data policies. Additionally, Gurret tracks information flow inside a filesystem container to prevent data leakage and policy violations. Our evaluations show that Gurret has minimal overhead for small to medium-sized files and that Gurret can support hundreds of custom metadata without losing transparency.

An emerging problem in systems security is controlling how a program uses the data it has access to. Information Flow Control (ifc) propagates restrictions on data by following the flow of information, for example if a secret value flows to a public value, that value should be considered secret as well. A common problem in ifc is reclassification of data, for instance to explicitly make data less restricted. An ifc mechanism often has strict flow rules in its normal operation, but reclassification by definition need to bypass these restrictions. This thesis proposes correctness criteria that aim to provide stronger semantic guarantees for the behavior of reclassification functions. We first conduct a survey on prior work in IFC, which concludes that little emphasis has been put on crystallizing such criteria. We then define a set of criteria for reclassification and implement a parser to enforce these criteria. If a piece of code is successfully analyzed by the parser, then that code can be safely used to reclassify data. Rust is emerging as one of the more prominent languages for systems programming due to its memory safety, and we conjecture this can be analogously continued to target ifc as well.

Elite football today is infused with technologies for collecting data on players' performance and health. Technologies such as wearable health and fitness trackers, full-body medical scans, and positional systems mounted around a stadium give coaches a wide range of accurate data about their players. However, coaches do not have the time or ability to analyze all this data for each player to give them individualized training schedules. Hence, coaches need tools to collect, analyze, summarize, and present the data to them in a much more consumable format. Existing systems within sports technology are isolated, only collecting their own data and using it for their own purposes. Hence, Sáhka will break out of this norm and create a novel system within this domain. Sáhka is a system that federates relevant data from several different sources. The data is processed, stored, analyzed, and presented visually to the users. Additionally, by collecting a large quantity and variety of data, Sáhka acts as a Big Data repository for real-time sports data. In the future, Sáhka will be used as a platform for training Machine Learning algorithms.

There is a strong culture for using safety risk management tools to monitor the different parts of the operation in the aviation industry. However, most of the monitoring that is being done is done on the technical conditions of the aircraft rather than on the pilot. While there is an expectation that the pilots self-assess their health regarding how fit the pilot is to operate an aircraft, all the tools given are checklist-based. Checklists are widely used in the aviation industry to ensure all tasks are done as they should for mechanics and pilots. However, the drawback of checklist-based systems is that they do not monitor anything over time. As a pilot has responsibility for many passengers on every flight, the consequences of mistakes can be considerable. By not monitoring the health over time, some of the crucial information when considering whether the pilot is fit to fly or not may be forgotten. Fatigue and stress are two essential topics for ensuring the focus is on operating the aircraft rather than either zoning out or being concerned about something else during flight. As the EU work hour regulations exempt everyone within the aviation industry, pilots can work at any time during the day. As the pilots can work at any given point during the day, they have to self-regulate whether they can work. If they do not track topics such as sleep and nutrition, they can be fatigued and lose focus on the work to be done. This thesis presents Gearggus, a self-assessment tool that can assist the pilot in assessing their health based on the information given by a questionnaire. The user answers questions based on how important the data is monitored over time. Based on the answers, there is calculated feedback on how ready the pilot is to operate an aircraft. The data is presented on a history page, so the user can see what the score is based on and how to adjust to gain a better score. Gearggus was evaluated with a qualitative interview with experienced aviation personnel and the Department of Aviation employees at the Unversity of Tromsø. Both parties acknowledge the issues Gearggus is trying to solve, but with modifications to the system. The required changes differ between the parties.

Hearing loss, and assistive technologies to compensate for the loss, are becoming more and more regular. Hearing aids have improved the quality of life for many suffering from hearing loss but are still insufficient in some social settings. The café problem rises when there are a group of people talking in a relatively noisy environment where one person has hearing aids. Even with modern advancements, such as speech recognition and noise cancellation, people using hearing aids have difficulties differentiating the group's conversation from other noises. This thesis will provide the architecture, design, implementation and evaluation of a mobile application as a first step in creating a system that can counter this café problem. A critical factor in a system like this is to reduce the audio latency to a minimum. We investigate where latency is introduced in the system by creating an experimental setup and evaluating the system. We implement a prototype system and use the experimental setup to identify latency-inducing components. We discuss how this latency can be reduced and bring forward future steps that must be made in completing the system.

This thesis aims to investigate the role of edge computing in a smart nudging system. A smart nudging system has requirements for efficient data processing of personal and context-aware data from heterogeneous sources. Furthermore, a smart nudging system needs to protect and preserve the privacy of data within the system. Edge computing has been proposed as a computing paradigm in a smart nudging system to accommodate some of these requirements. The edge computing paradigm makes promises of low latency, context-aware data collection and contributions to privacy when running on an edge device. However, edge computing has limitations in resources for heavy computations and storage. Therefore, a smart nudging system, NuEdge, has been proposed to utilize edge computing resources integrated with cloud computing, a local server, and IoT devices for better performance, privacy storage, and data off-loading. Further, a prototype of the NuEdge system has been implemented to discover the possibilities and limitations of the prototype in a real-world scenario. The primary nudge goal of the system is to improve physical activity for inactive users. By gathering research on edge computing and smart nudging, combined with the implementation's observations, has edge computing's role in smart nudging been evaluated. Edge computing has significantly contributed to efficient data collection in a smart nudging system and lower latency for data transmissions. Future work should include a large-scale prototype and new technologies like 5G to investigate the limitations of edge device capabilities such as power consumption, storage, and computational power.

The edge computing paradigm has recently started to gain a lot of momentum. The field of Artificial Intelligence (AI) has also grown in recent years, and there is currently ongoing research that investigates how AI can be applied to numerous of different fields. This includes the edge computing domain. In Norway, there is currently ongoing research being conducted that investigates how the confluence between AI and edge computing can be used to hinder fish crime, by stationing surveillance equipment aboard fishing vessels, and perform all the monitoring directly on the vessel with support of AI. This is challenging for several reasons. First and foremost, the equipment needs to be stationed on the vessel, where actors may impose a threat to it and attempt to damage it, or interfere with the analytical process. The second challenge is to enable multiple machine learning pipelines to be executed effectively on the equipment. This requires a versatile computation model, where data is handled in a privacy preserving manner. This thesis presents Áika, a distributed edge computing system that supports machine learning inference in such untrusted edge environments. Áika is designed as a hierarchical fault tolerant system that supports a directed acyclic graph (DAG) computation model for executing machine inference on the edge, where a monitor residing in a trusted location can ensure that the system is running as expected. The experiment results demonstrate that Áika can tolerate failures while remaining operable with a stable throughput, although this will depend on the specific configuration and what computations that are implemented. The results also demonstrate that Áika can be used for both simple tasks, like counting words in a textual document, and for more complex tasks, like performing feature extraction using pre-trained deep learning models that are distributed across different workers. With Áika, application developers can develop fault tolerant and different distributed DAGs composed of multiple pipelines.

The Norwegian healthcare sector is vast, with a substantial number of organizations employing plenty of people, and thereof, 7,300 are customers of Norsk Helsenett. Maintaining and keeping up-to-date information about access rights for these customers is a difficult, time-consuming, and manual task, especially since organizations often change personnel without notifying the system administrators. The task is worsened since the access rights are often tied to an account using username and password as login credentials, which are easy to forget in the midst of many other login credentials. This thesis proposes a proof of concept system for solving these problems. The proof of concept system utilizes HelseID, a unified login solution designed for health care personnel, offering authentication through known identity providers. Further, Altinn, a trusted third-party state enterprise, offers an API for service owners to extract information about organizational roles. The system acquires role assignments data from Altinn and delegates it to users when they are authenticated. The main goal is to automatically and dynamically keep information about access rights up-to-date, alleviate the task from system administrators, and make login simpler for the end-user by using a login solution they are familiar with.

Skin cancer is one of the most common types of cancer and is usually classified as either non-melanoma and melanoma skin cancer. Melanoma skin cancer accounts for about half of all skin cancer-related deaths. The 5-year survival rate is 99% when the cancer is detected early but drops to 25% once it becomes metastatic. In other words, the key to preventing death is early detection. Foraminifera are microscopic single-celled organisms that exist in marine environments and are classified as living a benthic or planktic lifestyle. In total, roughly 50,000 species are known to have existed, of which about 9,000 are still living today. Foraminifera are important proxies for reconstructing past ocean and climate conditions and as bio-indicators of anthropogenic pollution. Since the 1800s, the identification and counting of foraminifera have been performed manually. The process is resource-intensive. In this dissertation, we leverage recent advances in computer vision, driven by breakthroughs in deep learning methodologies and scale-space theory, to make progress towards both early detection of melanoma skin cancer and automation of the identification and counting of microscopic foraminifera. First, we investigate the use of hyperspectral images in skin cancer detection by performing a critical review of relevant, peer-reviewed research. Second, we present a novel scale-space methodology for detecting changes in hyperspectral images. Third, we develop a deep learning model for classifying microscopic foraminifera. Finally, we present a deep learning model for instance segmentation of microscopic foraminifera. The works presented in this dissertation are valuable contributions in the fields of biomedicine and geoscience, more specifically, towards the challenges of early detection of melanoma skin cancer and automation of the identification, counting, and picking of microscopic foraminifera.

In this project we developed a web portal for students and by students in order to be able to help each other understanding coding in the different subjects across fields of study. The web portal is implemented using HTML programming language and its templates, and running the front-end with flask web application framework. A database has been made for collecting and storing code snippets.

This project aims to explore creating a relational database of Conflict-Free Replicated Relations (CRR) that is generally applicable for most user applications. We aim to achieve generality by using SQLite because it is a commonly used embedded Database Management System that comes default in all Android or iOS devices. The implementation challenge is making CRR using only SQL. We propose a two-layer database design using Causal Lenght-set for rows in the relations and Last Write Wins-registers for attributes in the rows. Achieving this means implementing methods to propagate data between layers, storing necessary metadata that automatically upholds CRDT functionality. Merges are performed entirely in SQLite. CL-set functionality decides the state of each row, and LWW-registers determine the values of each attribute. In merges, the SQL script resolves reference integrity constraints. We base experiments on data and client-centric consistency and availability as latency. The experiments highlight the strong eventual consistency across two merge methods. Full merges are when the entire system merges periodically, which grants the system complete consistency on demand. Partial merges simulate that even though nodes are not always available, the system can uphold a consistency threshold. Availability is a trade-off compared to a local datastore when measured as transaction latency. However, the system's scalability outperforms other fairer comparisons like Paxos and Raft.

Through time, we as a society have been tested with infectious disease outbreaks of different magnitude, which often pose major public health challenges. To mitigate the challenges, research endeavors have been focused on early detection mechanisms through identifying potential data sources, mode of data collection and transmission, case and outbreak detection methods. Driven by the ubiquitous nature of smartphones and wearables, the current endeavor is targeted towards individualizing the surveillance effort through a personalized health model, where the case detection is realized by exploiting self-collected physiological data from wearables and smartphones. This dissertation aims to demonstrate the concept of a personalized health model as a case detector for outbreak detection by utilizing self-recorded data from people with type 1 diabetes. The results have shown that infection onset triggers substantial deviations, i.e. prolonged hyperglycemia regardless of higher insulin injections and fewer carbohydrate consumptions. Per the findings, key parameters such as blood glucose level, insulin, carbohydrate, and insulin-to-carbohydrate ratio are found to carry high discriminative power. A personalized health model devised based on a one-class classifier and unsupervised method using selected parameters achieved promising detection performance. Experimental results show the superior performance of the one-class classifier and, models such as one-class support vector machine, k-nearest neighbor and, k-means achieved better performance. Further, the result also revealed the effect of input parameters, data granularity, and sample sizes on model performances. The presented results have practical significance for understanding the effect of infection episodes amongst people with type 1 diabetes, and the potential of a personalized health model in outbreak detection settings. The added benefit of the personalized health model concept introduced in this dissertation lies in its usefulness beyond the surveillance purpose, i.e. to devise decision support tools and learning platforms for the patient to manage infection-induced crises.

This thesis will investigate different robotic manipulation and grasping approaches. It will present an overview of robotic simulation environments, and offer an evaluation of PyBullet, CoppeliaSim, and Gazebo, comparing various features. The thesis further presents a background for current approaches to robotic manipulation and grasping by describing how the robotic movement and grasping can be organized. State-of-the-Art approaches for learning robotic grasping, both using supervised methods and reinforcement learning methods are presented. Two set of experiments will be conducted in PyBullet, illustrating how Deep Reinforcement Learning methods could be applied to train a 7 degrees of freedom robotic arm to grasp objects.

This thesis discusses the application of optimizations to machine learning algorithms. In particular, we look at implementing these algorithms on specialized hardware, I.e. a Graphcore Intelligence Processing Unit, while also applying software optimizations that have been shown to improve performance of traditional workloads on general purpose CPUs. We discuss the feasibility of using these techniques when performing Matrix Factorization using Stochastic Gradient Descent on an IPU. We implement a program doing this, and show the results of changing different parameters during the running of SGD. We demonstrate that while machine learning is inherently approximate this does not mean that all approximate computation techniques are applicable, and that indeed some of these techniques require a more measurable level of approximation that is given by there being a correct answer, I.e. that the algorithm being approximated is not inherently approximate from the start. We also show that other techniques can be applied to reduce the time it takes for SGD to converge.

Virtual reality (VR) technology is a promising tool in physical rehabilitation. Research indicates that VR-supported rehabilitation is beneficial for task-specific training, multi-sensory feedback, diversified rehabilitation tasks, and patient motivation. Our first goal was to create a biomechatronics laboratory with a VR setup for increasing immersion and a motion platform to provide realistic feedback to patients. The second goal was to investigate possibilities to replicate features of the biomechatronics laboratory in a home-based training system using commercially available components. The laboratory comprises of a motion platform with 6-degrees-of-freedom (Rexroth eMotion), fitted with a load cell integrated treadmill, and an Oculus Quest virtual reality headset. The load cells provide input for data collection, as well as VR motion control. The home-based rehabilitation system consists of a Nintendo Wii Balance Board and an Oculus Rift virtual reality headset. User studies in the laboratory and home environment used direct observation techniques and self-reported attitudinal research methods to assess the solution’s usability and user experience. The findings indicate that the proposed VR solution is feasible. Participants using the home-based system experienced more cybersickness and imbalance compared to those using the biomechatronics laboratory solution. Future studies will look at a setup that is safe for first patient studies, and exercises to improve diagnosis of patients and progress during rehabilitation.

People with diabetes live a life of constant need to monitor and manage their illness. Unfortunately, a large number of people with diabetes suffer from diabetes-related distress and depression. A better diabetes consultation can directly impact the health of people with diabetes, both mentally and physically. In this thesis, healthcare practices that previous studies have reported lead to a successful diabetes consultation with an improved health outcome are identified through literature review. The findings are used as the inspiration for designing and implementing a dynamic electronic patient-reported outcome measures (ePROM) mobile application. The application promotes the identified healthcare practices by using the patient-gathered data of each user and tailoring the contents of their questionnaire to them. The application is usability tested with the System Usability Scale is deemed to have good usability (83.75/100) and be a facilitator for better consultations.

\textit{Diabetes} is a disease defined by raised blood glucose levels, and people with the disorder have to adhere to a strict self-management regime to avoid short- and long-term complications. Today the World Health Organization reports that over 420 million people have diabetes, and the condition is one of the most common causes of death worldwide. This thesis presents the design and implementation of a group-based motivational application for people with diabetes. The proposed application applies persuasive design techniques and aspects of motivational theory to help people stay motivated in their management of the disorder through groups. Thus, decreasing the risk for short- and long-term complications. The result is \textit{Salutem}, an application where users can define custom goals, participate in groups, and receive progression updates and notifications that encourage them to motivate each other. The thesis also shows how other applications can arise from openness with user data. For example, with other application's integration to modern technologies such as HealthKit, the proposed application gathers blood glucose data from the user's CGM through HealthKit. Because of the ongoing pandemic, an extensive user study of the implemented system is not conducted. As a result, it could not be determined if the application will affect the user's motivation. However, a user study of the application's usability is conducted on a small group of users who reported that the application had an appealing design and work as intended.

Explorative data visualization is a widespread tool for gaining insights from datasets. Investigating data in linked visualizations lets users explore potential relationships in their data at will. Furthermore, this type of analysis does not require any technical knowledge, widening the userbase from developers to anyone. Implementing explorative data visualizations in web browsers makes data analysis accessible to anyone with a PC. In addition to accessibility, the available types of visualizations and their interactive latency are essential for the utility of data exploration. Available visualizations limit the number of datasets eligible for use in the application, and latency limits how much exploring the users are willing to do. Existing solutions often do all the computation involved in either the client application or on a backend server. However, using the client limits performance and data size since hardware resources in web browsers are scarce, and sending large datasets over a network is not feasible. Whereas server-based computation often comes with high requirements for server hardware and is limited by network latency and bandwidth on each interaction. This thesis presents Slicer, a framework for creating explorative data visualizations in web browsers. Applications can be created with minimal developer effort, requiring only a description of the visualizations. Slicer implements bar charts and choropleth maps. The visualizations are linked and can be filtered either by brushing or clicking on single targets. To overcome the hurdles of pure client- and server-reliant solutions, Slicer uses a hybrid approach, where prioritized interactions are handled client-side. Recognizing that different types of interactions have different latency thresholds, we trade the cost of switching views for low latency on filtering. To achieve real-time filtering performance, we follow the principle that the chosen resolution of the visualizations, not data size, should limit interactive scalability. We describe use of data tiles accommodating more interactions than shown in earlier work, using an approach based on delta differencing, which ensures constant time complexity when filtering. For computing data tiles, we present techniques for efficient computation on consumer hardware. Our results show that Slicer can offer real-time interactivity on latency-sensitive interactions regardless of data size, averaging above 150Hz on a consumer laptop. For less sensitive interactions, acceptable latency is shown for datasets with tens of millions of records, depending on the resolution of the visualizations.

Climate change is going to change what we know about the arctic tundra. Patterns in the behavior of the wildlife that lives there are predicted to undergo a shift, and it will therefore be important to have reliable sources of empirical data, so that we can understand how these developments are playing out. The arctic tundra is remote and difficult to deploy sensing instruments on, and signal coverage is unreliable. Finding a way to monitor them reliably from a distance is needed. This thesis describes how a prototype for a Wireless Sensor Network was designed, implemented, and tested, with the aim of connecting Observational Units together in a local cluster, and cooperate amongst themselves to propagate monitoring data to external servers. The system was designed so that nodes can dynamically discover neighboring nodes within their range, and gossip knowledge about where sinks are in the network. Sinks are nodes which have managed to establish a link with an external server, and the paths to these sinks are spread across the network. Such that if only node in the entire cluster is a sink, then data from every node has a path outside of the cluster. Results from running validation shows that the implemented prototype func- tions as intended, but experiments have revealed apparent weaknesses. The number of paths which are shared in gossiping shows an exponential growth when the number of nodes in a cluster grows linearly. The experiments into bundling and monitoring-data propagation shows that combining data together causes a reduction in these types of transmissions by a factor equal to that of the number of data fragments which are combined, however the Partial Bundle Policy measure to increase throughput for fringe nodes has unexpected consequences. The prototype system works as intended per the design. We have found however that the system is not scalable due to the extent of the accumulated path knowledge. Suggestions for avenues to address this has been outlined in the discussion chapter. There is a need to explore how something similar to this prototype would look and perform in a real-life deployment on the arctic tundra.

In this thesis, the ability of CharRNN models learning to compose guitar music using varying representations of guitar tablature is explored. I utilize a well-versed sequential model of LSTM cells, and investigate the ability of said model to input, and predict both character to character, and sequence to sequence, following the principles of natural language processing and music information retrieval. The study was conducted on datasets consisting of data naïvely retrieved from a subset of classical guitar tablature. With regards to tablature structure, the experiments uncover a clearly superior form for character to character prediction, producing a model capable of composing seemingly musically coherent phrases. The work is not fully able to compare the character predictor with the sequence predictor and further details how this could potentially be alleviated.

People with intellectual disabilities have difficulties in reaching the World Health Organization's (WHO) suggested level of physical activity. Previous research shows that participating in physical activities often is related to self-efficacy in a physical activity setting and personal motivation. As physical activity has significant effects on physical and mental health, this thesis aimed to develop a mobile application that could help people with intellectual disabilities be more physically active. In the process of creating an encouraging and user-friendly mobile application, this project includes literature reviews, meetings with experts in the field, discussions with special education teachers and teachers working with people with intellectual disability, and the author's own experience with this user group. The project relies on guidelines and theory to create a user interface to fit people with intellectual disabilities. This thesis presents a cross-platform mobile application that combines the digital and real world. Using augmented reality, players walk around in the real world looking after digital garbage. As they walk, they will find garbage on the ground and get the option of sorting the garbage in the correct garbage bins. The game's main objective is to look for garbage and throw it in the correct garbage bin. As users progress throughout the game, they earn stars based on their step count while playing the game. Together with family members or assistants, they can add weekly physical activity goals and earn special rewards created by family members or assistants. Usability testing is mainly done on special education teachers, social workers, psychologist, and researchers working with people with intellectual disabilities. It revealed that creating a mobile application focusing on everyday life scenarios can have a potential value for the targeted user group. However, testing also showed that using augmented reality can be challenging. Long-term testing on individuals with an intellectual disability will start in the upcoming weeks in a study conducted by the University Hospital of North Norway (UNN), in collaboration with UiT The Arctic University of Norway.

The arctic tundra is one of the ecosystems that is most affected by climate changes. The effects of these changes on the wildlife in the arctic are therefore critical to monitor. To monitor the changes, small computing devices with sensors and cameras, known as Observational Units, can be used. Using a cluster network of interconnected observational units, so that data can be reported from the most distant nodes to a homebase, introduces problems as the node’s local clocks tend to skew away from each other based on the environment they are in. This thesis aims to address the problem of clock synchronization in cluster networks that are disconnected from constant power and the internet. This thesis describes how we designed, built, and tested a prototype software solution for an interconnected Wireless Sensor Network using observational units built for arctic climates. We designed and built a two-phased system where the nodes dynamically join a network and synchronize their duty cycling and clocks with each other, creating a synchronized cluster of nodes that allows for data to be propagated, where any of the nodes can become sink nodes if they have a connection to a homebase. The nodes use a built-in clock synchronization operation to achieve synchronized clocks across the cluster network. The prototype system can perform the operations as planned. However, the results show that the system scales poorly when introducing new sinks to the network while running in the operational phase, as the paths shared scales exponentially, in contrast to when a sink is introduced in the starting phase of the system, where the transmissions grow linearly with a degree based on the number of nodes in the network. The time synchronization experiments also showed that the network is able to remain synchronized, although transmission number is a concern when the network does not have any sink nodes.

This master's thesis reports on security of a Cyber-Physical System (CPS) in the department of industrial engineering at UiT campus Narvik. The CPS targets connecting distinctive robots in the laboratory in the department of industrial engineering. The ultimate objective of the department is to propose such a system for the industry. The thesis focuses on the network architecture of the CPS and the availability principle of security. This report states three research questions that are aimed to be answered. The questions are: what a secure CPS architecture for the purpose of the existing system is, how far the current state of system is from the defined secure architecture, and how to reach the proposed architecture. Among the three question, the first questions has absorbed the most attention of this project. The reason is that a secure and robust architecture would provide a touchstone that makes answering the second and third questions easier. In order to answer the questions, Cisco SAFE for IoT threat defense for manufacturing approach is chosen. The architectural approach of Cisco SAFE for IoT, with similarities to the Cisco SAFE for secure campus networks, provides a secure network architecture based on business flows/use cases and defining related security capabilities. This approach supplies examples of scenarios, business flows, and security capabilities that encouraged selecting it. It should be noted that Cisco suggests its proprietary technologies for security capabilities. According to the need of the project owners and the fact that allocating funds are not favorable for them, all the suggested security capabilities are intended to be open-source, replacing the costly Cisco-proprietary suggestions. Utilizing the approach and the computer networking fundamentals resulted in the proposed secure network architecture. The proposed architecture is used as a touchstone to evaluate the existing state of the CPS in the department of industrial engineering. Following that, the required security measures are presented to approach the system to the proposed architecture. Attempting to apply the method of Cisco SAFE, the identities using the system and their specific activities are presented as the business flow. Based on the defined business flow, the required security capabilities are selected. Finally, utilizing the provided examples of Cisco SAFE documentations, a complete network architecture is generated. The architecture consists of five zones that include the main components, security capabilities, and networking devices (such as switches and access points). Investigating the current state of the CPS and evaluating it by the proposed architecture and the computer networking fundamentals, helped identifying six important shortcomings. Developing on the noted shortcomings, and identification of open-source alternatives for the Cisco-proprietary technologies, nine security measures are proposed. The goal is to perform all the security measures. Thus, the implementations and solutions for each security measure is noted at the end of the presented results. The security measures that require purchasing a device were not considered in this project. The reasons for this decision are the time-consuming process of selecting an option among different alternatives, and the prior need for grasping the features of the network with the proposed security capabilities; features such as amount and type of traffic inside the network, and possible incidents detected using an Intrusion Detection Prevention System. The attempts to construct a secure cyber-physical system is an everlasting procedure. New threats, best practices, guidelines, and standards are introduced on a daily basis. Moreover, business needs could vary from time to time. Therefore, the selected security life-cycle is required and encouraged to be used in order to supply a robust lasting cyber-physical system.

Master's thesis in Business administration (BE501)

Master´s thesis in Information Systems (IS501)

In modern time, cloud services have been the go-to approach to store information and documents, and cloud makes documents accessible for users and make it easier to manage items. Cloud is an excellent option to have when data need to be available, but what if collaboration is only needed a few times a week? Once a month? Or unstable connection leading to disconnection to the network? A solution is a Local-first approach. This means all items and documents a user has created is store locally on the device it is created on and provide a feeling of ownership to what is created. Local documents are treated as primary copies, and potential backups stored in an external cloud own by big companies are secondary copies. Different communication technologies are explored, tested, and evaluated in this thesis in combination with a Local-fist approach and Automerge. A prototype is developed to connect all these ideas and concepts and to better understand which technologies do fit together and provide the necessary features to achieve a Local-first application for collaboration between computers and mobile.

The arctic tundra is observed to collect data to be used for climate research. Data can be collected by cyber-physical computers with sensors. However, the arctic tundra has a limited availability of energy. Consequently, the nodes rely on batteries and sleep most of the time to increase the battery-limited operational lifetime. In addition, only a few nodes can expect to be in reach of a back-haul wireless data network. Consequently, the nodes have on-node wireless local area networks to reach nearby neighbor nodes. To increase the availability for remote clients to the data collected by the nodes, a set of shadow nodes are used. These are always on, and always have access to a back-haul network. Data from an edge node on the arctic tundra propagates to the shadow nodes either directly over a back-haul network, or via a neighbor node with a back-haul network. The purpose is to make the data produced by an edge node available to a client even when the edge node sleeps or no network access is available. A statistical analysis is done to characterize the prototype’s behavior under a set of edge-node behaviors. To validate the statistical analysis a prototype system is developed and used in a set of performance-measuring experiments. Experiments are done with 10 to 1,000,000 nodes, different probabilities of nodes being awake, and different probabilities of the back-haul network being available. Edge and shadow nodes are emulated as Go functions and executed on a high-performance computer with thousands of cores. Different wireless networks are emulated albeit in a simplified way. A run-time simulation system is developed to control the prototype and conduct the experiments. The results for the prototype show that if the single synchronization chance is low or the desired time to get the latest data should be minimized, an additional data delivery path should be considered on the edge node’s side. Synchronization via the right neighbor principle adds an extra communication channel which increases the data availability level by 50%-100%, but the resource demand grows by 30% per unit. The time required to get the latest data from edge nodes decreases by a factor of 1.75. The results for the simulation show that the cumulative network throughput of approximately ≈ 2100 MB/s and the Generated Data Amount ≈ 25000 MB/s can be achieved at the cost of ≈ 80 KB RAM per emulated node. The results show that the statistical analysis and the results from the prototype as used by the simulation system match, but the statistical expectation considers a limited range of factors. Statistically derived values can be used as the input for the simulation, where they would be adjusted to get a more comprehensive result. The conclusions are that the Mask provides instant access to data storage for edge nodes. The Mask is fronted to clients which become able to retrieve the data asynchronously, even when edge nodes are offline.

Biological data is often visualized using networks. However, these networks face problems such as information overload, high interconnectivity, and high dimensionality. Existing approaches try to solve these problems by reducing the interactivity in favor of presenting more information or by using expensive hardware. This thesis aims to solve them using Virtual Reality (VR) and the Oculus Quest, an affordable VR headset, by taking advantage of the rich interactivity that VR offers. In order to test our hypothesis that Virtual Reality can be advantageous in the visualization of large biological networks, we built GeneNet VR, an open-source prototype of a VR application for the Oculus Quest for the interactive visualization of large biological networks. As a case study, we used two gene networks from MIxT, a real application that uses a 2-dimensional network visualization. We evaluated the performance and scalability of GeneNet VR and we conducted in-depth semi-structured interviews with several research scientists to evaluate the usability of our approach. Our result shows that the performance of the interactions for network visualization on a machine, reaches the 72 FPS required by the Oculus’ performance guidelines and that GeneNet VR scales for our largest network with 2693 nodes. We also evaluated the performance of GeneNet VR on the Oculus Quest hardware, which also achieved 72 FPS. The Oculus Quest is therefore an affordable option for the visualization of large datasets. From the interviews, we learned that GeneNet VR is an innovative and interesting visualization tool for large biological networks and that is easy to use even for novice VR users. Thus, VR hardware like the Oculus Quest should be considered a competitive solution for visualization tools, as described in this thesis. GeneNet VR is open-source and can be accessed with the following link: https://github.com/kolibrid/GeneNet-VR. We created also a video to show the different interactions that we can do with GeneNet VR to explore large biological networks: https://youtu.be/N4QDZiZqVNY.

Increases in data production and growing demands for more computing power leads to the current trend of outsourcing data and computation services to cloud providers. With data breaches and cases of data misuse becoming increasingly common, there is a high demand for secure systems. This, however, conflicts with the current data trust models. A solution to this that is becoming more common is the use of Trusted Execution Environment (TEE), aimed at guaranteeing code and data integrity and confidentiality. However, it has been shown that TEEs such as Intel's Software Guard Extensions (SGX) are susceptible to several types of side-channel attacks where an adversary may gain information of the code and data within a secure environment, breaking the confidentiality property. There are some ways to counter this, such as using oblivious primitives to hide access patterns which may leak information, but these are inefficient and add performance overhead to computation. Another way to ensure data confidentiality while simultaneously retaining the ability to perform computations on the data is through the use of Fully Homomorphic Encryption (FHE). FHE allows computing on encrypted data, preserving confidentiality and allowing outsourced computations to untrusted parties such as cloud providers. However, this type of encryption is malleable and lacks integrity protection, making it susceptible to integrity breaches where an adversary could modify the data resulting in a corrupt or incorrect plaintext after decryption. This thesis implements a library for performing FHE in SGX, written in a memory-safe programming language to strengthen the internal safety of software in SGX and reduce its attack surface. We evaluate our design and show that one can feasibly combine these concepts while providing stronger security guarantees with a minimal development effort.

As modern computer hardware becomes more affordable, the application areas of Artificial Intelligence (AI) expand and become more feasible to implement efficiently. This aspect opens up opportunities where AI robots can be used to support and improve many areas of human lives. One such area is education, and this thesis project combines AI and education to evaluate whether the results from that union can provide effective education. As a solution, a personalized language education robot that teaches the basics of the Italian language was implemented. The robot collects the user’s movie preferences using verbal communication and uses that information during the study sessions to personalize the learning experience. After implementation, the system has been tested on actual students to evaluate its educational effectiveness. The students filled out questionnaires that asked them whether they thought the system was effective in teaching a language, whether they thought personalization was a motivating factor or not, and they also provided their feedback as to how to improve the system in the future. The results showed that the personalized language education robot was, for the most part, effective in teaching the basics of the Italian language. However, the feedback from the users highlighted some key improvements that may be necessary if such a system were to be utilized to cover a broader subject.

Patients increasingly collect lifestyle and health-related data thanks to the explosion of sensors and healthcare applications. Multiple studies have shown that this data can be useful during consultations, resulting in more customised medical services. Moreover, this data permits clinicians to gain an overview of the patients’ conditions and enables patients to gain knowledge about the consequences of their lifestyle choices for their health. The goal of this thesis is to introduce self-collected health data by patients with diabetes into electronic health record systems (EHRs) and medical consultations. To achieve this goal, this thesis proposes, firstly, to study the state-of-the-art usage of self-collected health data during consultations. The thesis then identifies the acceptance barriers perceived by clinicians, patients, healthcare institutions, and EHR vendors that limit or obstruct the usage of self-collected health data during consultations. Thereafter, the thesis describes the FullFlow clinical decision support system (CDSS) addressing the identified acceptance barriers and finishes by its assessment by clinicians. The identification of acceptance barriers and the design of the FullFlow CDSS followed a participatory design approach supported by semi-structured interviews and open discussions used during workshops and focus groups involving clinicians, EHR vendors, healthcare institutions and patients with diabetes. The implementation of the FullFlow CDSS followed a combination of Agile and waterfall methodologies supported by a test-driven development approach. The assessment of the FullFlow CDSS relied on a case-study. Summarising, despite multiple parties being interested in using self-collected health data in a medical context, its usage is still rare and locked into controlled environments, even if some cloud-based solutions permit the integration of this type of data into EHRs. The limited usage of this type of data in medical context can be explained by the numerous acceptance barriers perceived by clinicians, patients, EHR vendors and healthcare institutions. Lack of data reliability, investment costs, lack of practice and training and time consumption are examples of acceptance barriers. The FullFlow CDSS comprises three solutions to address these acceptance barriers: 1) a computer science model for extracting relevant health data and managing data reliability, using a hypothesis-and-test strategy; 2) a dashboard permitting clinicians to consult relevant information regarding patients’ conditions; and 3) an architecture facilitating the integration of patients’ self-collected health data with their health applications in EHRs. The assessment of this system by clinicians showed that this system can integrate self-collected health data during medical consultations and that it would be useful during clinicians’ daily consultations. The next step is to test and verify the developed system in real settings.

Recent advancements in human-robot interaction have led to tremendous improvement for humanoid robots but still lacks social acceptance among people. Though verbal communication is the primary means of human-robot interaction, non-verbal communication that is proven to be an integral part of the human interactions is not widely used in humanoid robots. This thesis aims to achieve human-robot interaction via non-verbal communication, especially using hand-signs. It presents a prototype system that simulates hand-signs recognition in the NAO humanoid robot, and further an online questionnaire is used to examine people's opinion on the use of non-verbal communication to interact with a humanoid robot. The positive results derived from the study indicates people's willingness to use non-verbal communication as a means to communicate with humanoid robots, thus encouraging robot designers to use non-verbal communications for enhancing human-robot interaction.