Browsing by Author "Hudson, Jonathan William"

Now showing 1 - 4 of 4
  • Thumbnail Image
    ItemOpen Access
    Driving Anomaly Detection Using Recurrent Neural Networks
    (2022-03-25) Sabour, Sepehr; Ghaderi, Majid; Stefanakis, Emmanuel; Hudson, Jonathan William
    Deep learning has changed many aspects of our lives in recent years. Every day, the improvements in artificial intelligence make computers more capable of doing our daily tasks. Traffic management has never been separated from these changes. Researchers have proposed many machine learning solutions to help traffic management centers monitor vehicles’ activities in the transportation networks. Driving anomaly detection refers to finding unexpected vehicles, situations and traffic flows in the transportation systems. Many research works have been conducted recently to address driving anomaly detection problem, however each of these solutions has drawbacks. This thesis suggests two innovative solutions for detecting anomalies in intelligent transportation systems using recurrent neural networks (RNNs). A brief introduction of driving anomaly detection techniques and RNNs is presented in the first part of the thesis. Then in the second part, two suggested solutions, DeepFlow and ThirdEye, are discussed. DeepFlow is a method to detect abnormal traffic flows in smart cities. It is argued in this thesis that finding a complete dataset of vehicles’ behaviors in driving scenarios is very difficult. To address this issue the DeepFlow solution from this thesis applies machine learning techniques to reduce the requirement for a comprehensive dataset without loosing accuracy. ThirdEye, the second solution introduced, focuses on detecting anomalous behaviors of driver-less vehicles. This model works based on predicting the vehicle’s state in the future. By measuring the distance between the actual state of the vehicle and the predicted one, the system can detect more than 90% of anomalies. Three different recurrent neural networks were tested to determine the best for ThirdEye.
  • Thumbnail Image
    ItemOpen Access
    Evaluating the Emergent Effects of (Multiple) Security Mechanisms via Evolutionary Algorithms
    (2018-11-30) Hudson, Jonathan William; Denzinger, Jörg; Williamson, Carey L.; Safavi-Naeini, Reyhaneh
    Security mechanisms provide protection against system penetration and exploitation by providing coverage for vulnerabilities. However, security mechanisms often have demanding operational requirements that necessitate access to system resources and control of monitoring points. At the same time, users have particular requirements from programs they install, how they interact with these programs, and what performance they expect from their computing system. These combined requirements create a selection problem where the user desires to balance security coverage, through a choice of security mechanism(s), with system performance and functionality. This problem is known as the Effective Security-in-Depth problem. First, this thesis introduces a genetic algorithm to enable an evolutionary search for interaction event sequences for the problem of Effective Security-in-Depth. This methodology required the development of a fitness function that integrated numerous system metrics while addressing the variance found in event sequence simulation and measurement. Next, the steps for effectively implementing this methodology as a software tool are described. Finally, this thesis introduces three processes to use the tool to select between single security mechanisms for different usage profiles, compare and contrast subsets of security mechanisms, and evaluate examples of emergent misbehaviour such as system failure. The initial experimental evaluation validates the ability of the search for interaction event sequences to make progress despite the challenges of stochastic system measurement. The remaining experimental evaluations demonstrate the success of an application of each of the three processes. The evaluation supports that the developed method, tool, and processes are a viable solution to the problem of Effective Security-in-Depth.
  • Thumbnail Image
    ItemOpen Access
    Flow Size Prediction With Short Time Gaps
    (2024-06-03) Hosseini, Seyed Morteza; Ghaderi, Majid; Wang, Mea; Hudson, Jonathan William
    Having a priori knowledge about network flow sizes is invaluable in network traffic control. Previous efforts on estimating flow sizes have focused on long flows, where each flow is identified by a large time gap in the sequence of packets. However, many network control mechanisms such as load balancing and rate control achieve better performance when operating over flowlets, short flows that are separated by small time gaps in the sequence of packets. In this work, using extensive measurements, we investigate the feasibility of predicting the size of short flows, where the flow duration can be in the order of microseconds. Specifically, we deploy several popular workloads in a public cloud testbed, and collect both network and host traces for each workload. The network trace contains standard packet metadata, while the host trace contains high-level host statistics (e.g.,memory usage and disk I/O) and low-level function call traces (e.g.,malloc(), send()) that are captured during the execution of each workload via host instrumentation using eBPF. These traces are then used to train machine learning models for flow size prediction with varying time gaps ranging from microseconds to milliseconds. Our results indicate that: (1) It is feasible to predict short flow sizes with high accuracy, i.e., percentage error in 0-12% range, (2) the low-level traces lead to 10-20% improvement in prediction accuracy compared to using the network and high-level traces.
  • Thumbnail Image
    ItemOpen Access