SoC Theses

Permanent URI for this collection


Recent Submissions

Now showing 1 - 5 of 11
  • Item
    Harnessing unlabeled data for improving generalization of deep learning methods
    (Wichita State University, 2023-07) Shanmugasundaram, Deepika; Rattani, Ajita
    Recent advancements in Deep Learning, Artificial Intelligence, and Computer Vision have reached a critical stage, enabling researchers to explore the automatic extraction of individual demographic traits, known as soft-biometrics. This research aims to leverage unlabeled data in predicting soft-biometric traits, such as gender and age, using deep learning models. The objective is to develop a model that can accurately classify these traits by utilizing semi-supervised methods that rely on a limited amount of labeled data and a vast amount of unlabeled data. While unlabeled data may initially seem devoid of crucial information, this thesis explores how it can be effectively used to enhance classification accuracy, especially in scenarios where labeled data is scarce. This study evaluated the accuracy of different image classification models on the Celeb-A and NIR-VIS datasets using co-training, mix-up procedure, knowledge distillation, and blind distillation techniques. The results showed that incorporating these methods led to improvements in accuracy across both datasets and various attributes such as gender classification and smiling classification. Exploring the combined use of different techniques and investigating their synergistic effects could lead to further accuracy improvements. Evaluating the models on larger and more diverse datasets, analyzing their generalization capabilities, optimizing hyperparameters and architectures, and applying the techniques to other computer vision tasks were also identified as areas for future research.
  • Item
    Distributed misbehavior detection in UAV flocks
    (Wichita State University, 2023-07) Aguida, Mohamed Anis; Monroy, Sergio A.Salinas
    Unmanned aerial vehicles have become increasingly popular in many applications such as remote surveillance, reconnaissance, and precision agriculture. Often multiple UAVs form a swarm and perform their operations in a distributed, coordinated fashion. A rogue UAV in the flock can negatively disrupt the expected behavior and may jeopardize the objective of the mission, leading other UAVs to make incorrect decisions or even crash. This work introduces GRIFFIN, a distributed and lightweight misbehavior detection framework for UAV flocks. GRIFFIN relies on readily available packet metadata (e.g., GPS coordinates) and signal characteristics (e.g., RSSI measurements) and detects malicious UAVs by employing a “majority voting” protocol. We show that GRIFFIN requires only three honest nodes for correct operations. We implement and evaluate GRIFFIN on (a) a realistic UAV simulator (ArduSim) and (b) a Raspberry Pi+Navio-based drone testbed. We find that GRIFFIN outputs 100% successful detection with zero false negatives as long as less than half of the UAVs in the flock are not compromised. Our implementation on the real UAV testbed shows that runtime overhead of GRIFFIN is minimal (i.e., it requires less than 1 MB of memory and consumes less than 1% of CPU) and computes operation within 2:5 ms.
  • Item
    Private neural network auctions for additive manufacturing
    (Wichita State University, 2023-05) Shukla, Amey; Monroy, Sergio A.Salinas
    Additive Manufacturing is changing the way we construct, deliver, and consume objects by allowing customers to quickly build custom objects on-demand and in locations near to them. Additive manufacturing operators need to optimally control the manufacturing process, monitor tasks in real-time and set the prices for their built objects. Moreover, operators need to protect private data related to the prices paid by customers and their purchased objects. Existing works focus on either optimal control, real-time monitoring, price setting, or privacy, missing the advantages of jointly addressing them. Existing works also require vast computational resources to accomplish only one of these tasks. To address these issues, we develop a differentially-private distributed neural network auction that optimally allocates manufacturing resources and sets prices in a way that maximizes the profit of the manufacturer. The auction protects the privacy of the customers’ bids. Moreover, to reduce the computing time, we design a parallel computation algorithm for the neural network that is executed by a cluster of edge computing devices. We evaluate the proposed neural network through extensive simulations. We observe that it can jointly perform the operators’ tasks while maintaining the privacy of the customers. Our simulations also show that the parallel algorithm reduces the execution time.
  • Item
    Semantic scene understanding for intelligent robotics
    (Wichita State University, 2023-05) Yan, Fujian; He, Hongsheng
    This dissertation focuses on improving robotic scene semantics understanding and developing a new human-robot interaction (HRI) interface based on augmented reality (AR). To achieve a deep scene understanding, the proposed scene semantics understanding method consists of three parts: object detection, object semantic comprehension, and feedback on robotic comprehension. The method analyzes detected objects’ category, function, property, and composition to enable robots to understand object semantics and reason relations between objects. Additionally, the dissertation proposes a method for an intelligent industrial robot to comprehend spatial constraints for model assembly. The proposed method uses an extended generative adversary network (GAN) with a 3D long short-term memory (LSTM) network to composite 3D point clouds from a single or a few multiple-depth scans. The spatial constraints of the segmented point clouds are identified by a neural-logic network that incorporates general knowledge of spatial constraints in terms of first-order logic. The proposed HRI interface superimposes robot-centered and human-centered reality on the working space to construct a mutual understanding environment. The interface enables humans to communicate with robots through speech and immersive touching, constructing mutual understanding through the user’s commands, localization and recognition of objects, object semantics, and augmented trajectory. The user’s vocal commands are interpreted to formal logic, and finger touching is detected and coordinated. Real-world experiments show the effectiveness of the proposed interface.
  • Item
    Methods to achieve t-closeness for privacy preserving data publishing
    (Wichita State University, 2023-05) Gowda, Vikas Thammanna; Bagai, Rajiv
    Privacy Preserving Data Publishing is an area of research focused on developing methods of anonymizing sensitive relational data such that it can be published without compromising the privacy of the individuals the data represents. The t-closeness technique is one of the most popular techniques for preserving individual privacy in data. It involves generalizing and suppressing some attributes of a given table, after partitioning the set of all records of that table into equivalence classes that satisfy a certain constraint. We present three methods for anonymizing datasets addressing the drawbacks of the existing methods. We present a new method to partition the set of records of a table into such equivalence classes. The rst method has several advantages over the existing methods for this task. The classes generated by our method are near-optimal, in that they satisfy the t-closeness constraint for even the \smallest" t value for which t-closeness is achievable and useful for the given table, thereby providing the highest amount of privacy. The second method anonymizes data with multiple sensitive attributes such that the privacy parameter t for each can be selected individually. Our method partitions the data into fragments and selects appropriate numbers of records from each fragment to create equivalence classes with sensitive attribute distributions that are guaranteed t-close. Our method can easily be generalized to an arbitrary number of sensitive attributes and to sensitive attributes with continuous domains. In the third method we present an algorithm for generating equivalence classes in the presence of multiple sensitive attributes. The equivalence classes generated by our method satisfy t-closeness for even the smallest t value for which t-closeness is achievable and useful for the given dataset, thereby providing the highest possible amount of privacy.