Swarm drones - efficient machine learning and informatics

In 2020, worldwide consumer drone unit shipment was 5 million, which is expected to be 9.6 million in 2030. This generates a global drone market with 26.3 billion USD in revenue. The popularity of drones in civilian and professional environments has changed the way humans live and work. However, they have also brought new challenges and threats to the environment and society (e.g., property and personal damage caused by inappropriate drone operations or hostile drones) which requires more robust regulatory mechanisms and advanced technologies of drones. Supervision of tiny shape, high-mobility drones by humans is inefficient and inaccurate, whereas Artificial Intelligence (AI) methods, especially deep learning (DL) show potential in drone detection, classification and tracking. However, as data-driven models, the performance of DL models is decided by the quality of data and model structure. At the same time, the structural complexity of black-box DL models affects their explainability and energy efficiency. These factors affect the willingness of people to trust DL models. Therefore, this thesis aims to analyze the impact of drone informatics on DL behaviour, and achieve more efficient training of high-trustworthiness DL models by efficient drone informatics. This will require research on DL explainability, trustworthiness and efficiency. The aforementioned researches are all interrelated and highly relevant to the data. In this thesis, firstly, explainable AI (XAI) and DL trust factors are reviewed. A theoretical DL trustworthiness metric Quality of Trust (QoT) and a lifelong AI trust- worthiness supervision protocol are proposed. Secondly, a novel partially explainable Gaussian-process-based neural network structure is proposed. Compared with conventional machine learning methods, it is more transparent and without any sacrifice in accuracy. Thirdly, a GAN-TDA method is proposed to analyze the learning efficiency of convolutional layers on drone images and guide the collection of new data. Collecting new data with direction could boost the DL model performance more efficiently in time and cost. Fourthly, a transistor operations (TOs) model is proposed to analyze the DL energy consumption scaling law to different model architectures and settings. Finally, a physical visual neural stealth drone canopy is designed with the hard-to-learn design features analyzed by GAN-TDA and painted with adversarial evasion features to escape DL drone detection and classification. The canopy design method is further extended to swarm drone scenarios. This thesis shows: 1) both model explainability and performance are related to DL trustworthiness, and need a trade-off according to the QoT of different tasks; 2) combining human-understandable efficient drone informatics and the understanding of DL energy scaling laws can find high-efficiency datasets and network structures, resulting in efficient DL models with high trustworthiness; 3) The above knowledge can be used to formulate attacks on drone-related DL models to reduce their trustworthiness.