• Adversarial Thresholding Semi-Bandits

      Anjum, Ashiq; Bagdasar, Ovidiu; Xue, Yong; Bower, Craig (University of Derby, 2020-12)
      The classical multi-armed bandit is one of the most common examples of sequential decision-making, either by trading-off between exploiting and exploring arms to maximise some payoff or purely exploring arms until the optimal arm is identified. In particular, a bandit player wanting to only pull arms with stochastic feedback exceeding a given threshold, has been studied extensively in a pure exploration context. However, numerous applications fail to be expressed, where a player wishes to balance the need to observe regions of an uncertain environment that are currently interesting (exploit) and checking if neglected regions have become interesting since last observed (explore). We introduce the adversarial thresholding semi-bandit problem: a non-stochastic bandit model, where a player wants to only pull (potentially several) arms with feedback meeting some threshold condition. Our main objective is to design algorithms that meet the requirements of the adversarial thresholding semi-bandit problem theoretically, empirically and algorithmically, for a given application. In other words, we want to develop a machine that learns to select options according to some threshold condition and adapts quickly if the feedback from selecting an option unexpectedly changes. This work has many real-world applications and is motivated by online detector control monitoring in high-energy physics experiments, on the Large Hadron Collider. We begin by describing the adversarial thresholding semi-bandit problem (ATSBP) in terms of a multi-armed bandit with multiple plays and extending the stochastic thresholding bandit problem to the adversarial setting. The adversarial thresholding exponentially-weighted exploration and exploitation with multiple plays algorithm (T-Exp3.M) and an algorithm combining label efficient prediction (LET-Exp3.M), are introduced that satisfy theoretical and computational Research specifications, but either perform poorly or fail completely under certain threshold conditions. To meet empirical performance requirements, we propose the dynamic label efficient adversarial thresholding exponentially-weighted exploration and exploitation with multiple plays algorithm (dLET-Exp3.M). Whilst computational requirements match those for T-Exp3.M, theoretical upper bounds on performance are proven to be worse. We also introduce an ATSBP algorithm (AliceBandit) that decomposes the action of pulling an arm into selection and observation decisions. Computational complexity and empirical performance under two different threshold conditions are significantly improved, compared with exponentially weighted adversarial thresholding semi-bandits. Theoretical upper bounds on performance are also significantly improved, for certain environments. In the latter part of this thesis, we address the challenge of efficiently monitoring multiple condition parameters in high-energy experimental physics. Due to the extreme conditions experienced in heavy-ion particle colliders, the power supply to any device exceeding safe operating parameters is automatically shut down or tripped, to preserve integrity and functionality of the device. Prior to recent upgrades, a device or channel trip would halt data-taking for the entire experiment. Post-trip recovery requires a costly procedure both in terms of expertise and data-taking time. After the completion of the current upgrading phase (scheduled for 2021), the detector will collect data continuously. In this new regime, a channel trip will result in only the affected components of the experiment being shut down. However, since the new upgraded experiment will enable data-taking to increase by a factor of 100, each trip will have a significant impact on the experiments ability to provide physicists with reliable data to analyse. We demonstrate that adversarial thresholding semi-bandits efficiently identify device channels either exceeding a fixed threshold or deviating by more than a prescribed range prior to a trip, extending the state-of-the-art in high-energy physics detector control.
    • High Performance Video Stream Analytics System for Object Detection and Classification

      Anjum, Ashiq; Yaseen, Muhammad Usman (University of DerbyCollege of Engineering and Technology, 2019-02-05)
      Due to the recent advances in cameras, cell phones and camcorders, particularly the resolution at which they can record an image/video, large amounts of data are generated daily. This video data is often so large that manually inspecting it for object detection and classification can be time consuming and error prone, thereby it requires automated analysis to extract useful information and meta-data. The automated analysis from video streams also comes with numerous challenges such as blur content and variation in illumination conditions and poses. We investigate an automated video analytics system in this thesis which takes into account the characteristics from both shallow and deep learning domains. We propose fusion of features from spatial frequency domain to perform highly accurate blur and illumination invariant object classification using deep learning networks. We also propose the tuning of hyper-parameters associated with the deep learning network through a mathematical model. The mathematical model used to support hyper-parameter tuning improved the performance of the proposed system during training. The outcomes of various hyper-parameters on system's performance are compared. The parameters that contribute towards the most optimal performance are selected for the video object classification. The proposed video analytics system has been demonstrated to process a large number of video streams and the underlying infrastructure is able to scale based on the number and size of the video stream(s) being processed. The extensive experimentation on publicly available image and video datasets reveal that the proposed system is significantly more accurate and scalable and can be used as a general purpose video analytics system.
    • A Trust Evaluation Framework in Vehicular Ad-Hoc Networks

      Adnane, Asma; Franqueira, Virginia N. L.; Anjum, Ashiq; Ahmad, Farhan (University of DerbyCollege of Engineering and Technology, 2019-03-11)
      Vehicular Ad-Hoc Networks (VANET) is a novel cutting-edge technology which provides connectivity to millions of vehicles around the world. It is the future of Intelligent Transportation System (ITS) and plays a significant role in the success of emerging smart cities and Internet of Things (IoT). VANET provides a unique platform for vehicles to intelligently exchange critical information, such as collision avoidance or steep-curve warnings. It is, therefore, paramount that this information remains reliable and authentic, i.e., originated from a legitimate and trusted vehicle. Due to sensitive nature of the messages in VANET, a secure, attack-free and trusted network is imperative for the propagation of reliable, accurate and authentic information. In case of VANET, ensuring such network is extremely difficult due to its large-scale and open nature, making it susceptible to diverse range of attacks including man-in-the-middle (MITM), replay, jamming and eavesdropping. Trust establishment among vehicles can increase network security by identifying dishonest vehicles and revoking messages with malicious content. For this purpose, several trust models (TMs) have been proposed but, currently, there is no effective way to compare how they would behave in practice under adversary conditions. Further, the proposed TMs are mostly context-dependent. Due to randomly distributed and highly mobile vehicles, context changes very frequently in VANET. Ideally the TMs should perform in every context of VANET. Therefore, it is important to have a common framework for the validation and evaluation of TMs. In this thesis, we proposed a novel Trust Evaluation And Management (TEAM) framework, which serves as a unique paradigm for the design, management and evaluation of TMs in various contexts and in presence of malicious vehicles. Our framework incorporates an asset-based threat model and ISO-based risk assessment for the identification of attacks against critical risks. TEAM has been built using VEINS, an open source simulation environment which incorporates SUMO traffic simulator and OMNET++ discrete event simulator. The framework created has been tested with the implementation of three types of TM (data-oriented, entity-oriented and hybrid) under four different contexts of VANET based on the mobility of both honest and malicious vehicles. Results indicate that TEAM is effective to simulate a wide range of TMs, where the efficiency is evaluated against different Quality of Service (QoS) and security-related criteria. Such framework may be instrumental for planning smart cities and for car manufacturers.