The increasing usage of neural networks forms a threat to the cyber security of the system that uses the network, and unfortunately adversaries use this vulnerability to attack the system with adversarial examples. As neural networks can have complex structures with tens of thousands of parameters, they are hard for humans to understand. Hence, existing white-box methods use very limited network information and most state-of-the-art methods are based on gradient descent. In this work we further investigate the inner workings of a neural network and consider using intermediate network information for the creation of adversarial examples. We show that neuron activation values can be distinguished by the class of the data point and contain meaningful information about the prediction of the network. Based on this information, we propose a new, gradient-free method for creating adversarial examples based on a genetic algorithm. By covering a larger part of the search space and manipulating the neuron activation values, our success rate exceeds most state-of-the-art methods, such as DeepFool and RFGSM. We also find that the trade-off between success rate and distance has a huge impact on the results of a method, wherefore we recommend to carefully balance this trade-off by formulating an optimization formula with a separate loss and distance component.

During this project we were asked to explore the possibilities and limitations of creating a virtual therapist that can help people who suffer from a sleeping disorder. In order to answer this question, the goal of the project was to make such a therapist ourselves.

The first two weeks of the project were spent doing research. We looked into cognitive behavioural therapy and existing eHealth solutions for people with sleeping disorders. We also conducted research into existing chat bots, platforms and languages. At the end of the research phase we concluded that machine learning was the best way to make the agent smart.

During the design phase we stumbled upon the fact that we could not get access to a corpus of conversations between therapists and patients discussing sleep problems. This meant that we could not use machine learning on this corpus, which greatly influenced how smart the agent was going to be. We had to rethink our design choices concerning the artificial intelligence and decided to use a different platform. The personalized artificial intelligence was incorporated in our web application, which communicates with the code via our server.

The testing of the product is based on unit tests, where all modules and functions are tested individually after every new addition. The user interface was tested by setting up use cases. User tests on the final product were also performed. This gave us insight in how people with no knowledge about the system would interact with the system.

The main purpose of this project was to explore the possibilities and limitations for creating a virtual therapy that treats insomnia. Even though the solution we created cannot yet be used as a standalone therapy, it shows a promising glimpse of the many possibilities to conduct virtual therapy. A project group with more time, money and help from a psychologist could greatly improve the project by buying or generating a corpus which can then be used for machine learning and by letting the psychologist model the conversation. The future of virtual therapy has yet to be discovered, but this project already provides a bright outlook.