No description available.

No description available.

A crucial issue in the design of aircraft components is the evaluation of a larger number of potential design alternatives. This evaluation involves too expensive procedures, consequently, it slows down the search for optimal design samples. As a result, scarce or small number of design samples with high dimensional parameter space and high non-linearity pose issues in learning of surrogate models. Furthermore, surrogate models have more issues in handling qualitative data (discrete) than in handling quantitative data (continuous). These issues bring the need for investigations of methods of surrogate modelling for the most effective use of available data.   The thesis goal is to support engineers in the early design phase of development of new aircraft engines, specifically, a component of the engine known as Turbine Rear Structure (TRS). For this, tree-based approaches are explored for surrogate modelling for the purpose of exploration of larger search spaces and for speeding up the evaluations of design alternatives. First, we have investigated the performance of tree models on the design concepts of TRS. Second, we have presented an approach to explore design space using tree models, Random Forests. This approach includes hyperparameter tuning, extraction of parameters importance and if-then rules from surrogate models for a better understanding of the design problem. With this presented approach, we have shown that the performance of tree models improved by hyperparameter tuning when using design concepts data of TRS. Third, we performed sensitivity analysis to study the thermal variations on TRS and hence support robust design using tree models. Furthermore, the performance of tree models has been evaluated on mathematical linear and non-linear functions. The results of this study have shown that tree models fit well on non-linear functions. Last, we have shown how tree models support integration of value and sustainability parameters data (quantitative and qualitative data) together with TRS design concepts data in order to assess these parameters impact on the product life cycle in the early design phase.

Electroencephalography (EEG) is a common technique used to read brain activity from an individual, and can be used for a wide range of applications, one example is during the rehab process of stroke victims. Loss of motor function is a common side effect of strokes, and the EEG signals can show if sufficient activation of the part of the brain related to the motor function that the patient is training has been achieved. Reading and understanding such data manually requires extensive training. This thesis proposes to use machine learning to automate the process of determining if sufficient activation has been achieved. The process consists of a Long Short Term Memory (LSTM) Autoencoder that trains to extract features of the EEG data to be used for classification using various machine learning classification methods. In order to answer the research questions: “How to extract features from EEG signals using Autoencoders?” “Which supervised machine learning algorithm identifies as the best classification based on the features generated by the Autoencoder?” The results show that the accuracy varies greatly from individual to individual, and that the number of features created by the Autoencoder for the classification algorithms to work with has a large impact on accuracy. The choice of classification algorithm played a role for the result as well, with Support Vector Machine (SVM) performing the best, but had less impact than the previously mentioned factors.

Understanding runtime behavior in a real-time system may be difficult, as tools such as performance profilers and powerful debugging tools like those present in traditional personal computers may be unavailable, and computational resources can be limited for recording the runtime. Few of today's published literature evaluates the usability of runtime visualizations, and therefore this study will present solutions on how to create understandable runtime visualization. This study further evaluates one solution, conducting usability tests and performance tests on the solution. The solution proved to be effective in helping users understand runtime execution during the usability tests. Recording the runtime proved to adversely affect performance on the tested systems during the performance tests. Multiple methods of recording runtime were tested which all affected the performance. This performance impact was concluded to be insignificant in over 90% of normal use-cases, but programs that loaded the system heavily was significant. Most programs with heavy loads will see a relatively small but noticeable performance impact, while in some extreme cases, the execution times can increase by over 30% while recording the runtime.

The constantly increasing amount of shared data worldwide demands a continuously improved understanding of current smartphone security vulnerabilities and limitations to ensure secure communication. Securing sensitive enterprise data on a Bring Your Own Device (BYOD) setup can be quite challenging. Allowing multiple applications to communicate through Inter-process Communication (IPC) in a shared environment can induce a wide range of security vulnerabilities if not implemented adequately. In this thesis, multiple different IPC mechanisms have been investigated and applied with respect to confidentiality, integrity, and availability (CIA-triad) of a system including an Android application and a server, to enable a secure Single Sign-On (SSO) solution. Relevant threats were identified that could highlight vulnerabilities related to the use of IPC mechanisms provided by the Android OS such as AIDL, Messenger, Content Provider, and Broadcast Receiver. A Proof-of-Concept (POC) system for each IPC mechanism was developed and implemented with targeted mitigation techniques (MT) and best practices to ensure a high level of conformity with the CIA-triad. Additionally, each IPC mechanism was evaluated through a set of functional tests, a Grey-box penetration testing approach, and a performance analysis of the execution time and the total Lines-of-Code (LOC) required. The results shows that there are indeed different ways of achieving secure communication on the Android OS and thereby enabling a secure SSO solution by ensuring the inclusion of related MTs to prevent critical security vulnerabilities. Also, the IPC mechanism with the highest performance in relation to execution time and LOC is shown to be AIDL.

The technology of mixed reality provides possibility to digitalize tasks whichpreviously are tedious and time consuming. In this paper we have a digitalversion of following an instruction manual with the use of a headset that providescapability of displaying the steps of the manual in a virtual window.This headset further provides additional capabilities for us by providing dataand information of the user behaviour. More precisely we are looking intothe capability to detect confusion in user behaviour of users that are followinga set of instructions, using a mixed reality headset.The headset we are using is a HoloLens 2 which has multiple sensors,cameras, and other features that can provide a lot of information about theuser. Eye tracking is one of the features available, and it is the main featurewe focus on to analyze user behaviour. Eye fixations can provide a lot ofinformation about the behaviour and we are using this data, including otherdata, to develop a machine learning model to detect confusion.We look into if it is possible to observe when a user becomes confusedin a mixed reality environment. We also look into what is possible to retainuser immersion and actually handling confusion in a user.We found that a random forest classifier can classify the type of data wehave provided very well. An experiment with a manual that is designed tocause confusion is performed to collect data for the machine learning model.Even though only a few participants took part of the experiment, we canshow promising results with the confusion classifier.The result shows promising methods for detecting confusion and thenalso managing a confused user in a mixed reality learning environment.i

Bee colonies are heavily threatened by a parasite. An anonymouscompany solves the problem by detecting the parasites with a mobile appusing Machine Learning (ML). It is based on cloud based computations andthey are currently exploring if it is possible to do the computations on thedevice instead. The recent advances in ML, object detection and mobile deviceshave made it possible to use very efficient models for detection. SSD MobileNetV2FPNLite is a lightweight model used in this thesis and it has supportfor TensorFlow and the TFLite format. It is trained and tested on a provideddataset and achieves relatively high mean Average Precision (mAP), small sizeand fast inference time. In conclusion the model shows that on device executionof object detection is possible and viable. Further improvements of overallmodel performance and video stream inference are relevant topics for thefuture.Supervisor:

ABSTRACT  Cyberattacks are constantly evolving and phishing activities have risen steeply in the last few years. As the number of online users is increasing so as the phishing attacks and scams are increasing too. It is even more surprising in the presence of the most sophisticated technical security measures and online users are continually becoming the victim of phishing attacks that causing financial and emotional loss. Phishing attacks involve deceiving a target user into revealing their most important personal information such as ID, password, username, bank card, or other sensitive information to the cybercriminals. The typical way to instigate a phishing attack by sending malicious emails that may contain malware or a link to a phishing website. It is evident from various phishing reports that despite the most sophisticated and expensive technical security measures, the phishing attacks are proved to be still successful. This is happening because phishing techniques bypass technical security measures and try to exploit vulnerabilities associated with human and use social engineering to reach its target. Therefore, in this situation, anti-phishing awareness is the most effective tool that can protect internet users against phishing attacks. Anti-phishing awareness material can be delivered in a number of methods; however, the effectiveness of these awareness delivery methods is an open question among the researcher community and the anti-phishing awareness program designers. Which method is more effective in anti-phishing awareness-raising, increasing overall users’ confidence in dealing with phishing emails, and which method users preferred more? In an attempt to address all these questions, we conducted experimental research involving online users with different demographic backgrounds. We design and deliver and online anti-phishing awareness-raising material in three formats, video-based, text-based, and infographic-based. We found all training methods significantly improve the accuracy rate of identifying phishing and genuine emails. The training decreased the false-negative rate and also reduced the false positive rate among the participants of all training groups when compared with a control group. However, our study did not find one awareness delivery method significantly more effective than other methods in transferring knowledge. However, the study found video and infographic methods as most preferred by the users. This study also found an interesting result that the difference between the accuracy of identifying phishing emails of participants who received training in their preferred learning method and the accuracy of participants who received training in other methods was not significantly different. These results serve researchers, students, organizations, cybersecurity expert, and security awareness program designers, who are interested in understating the relationship between different awareness rising delivery methods and their effectiveness in educating internet users about prevention from phishing attacks.

Existing supervised learning-based detectors for precision agriculturehave previously achieved high accuracy in challenging classificationtasks. However, their performance deteriorate when presented with new environmentsdue to variations in observed objects and surrounding environment.Accordingly, it is desired to accelerate a detector’s adaptability when operatingon new environments. Therefore, this thesis proposes an effective methodfor semi-supervised object detection that can adapt detectors to new environmentswith minimal manual labeling effort. Experimental results show thatthe proposed method reduces annotation efforts by more than 400x while attainingsimilar accuracy to supervised learning alternatives.