All students and faculty are welcome to attend the final defense of EECS graduate students completing their M.S. or Ph.D. degrees. Defense notices for M.S./Ph.D. presentations for this year and several previous years are listed below in reverse chronological order.
Students who are nearing the completion of their M.S./Ph.D. research should schedule their final defenses through the EECS graduate office at least THREE WEEKS PRIOR to their presentation date so that there is time to complete the degree requirements check, and post the presentation announcement online.
UPCOMING DEFENSE NOTICES
When & Where:
May 7, 2021 - 11:00 AMZoom Meeting, please contact jgrisafe@ku.edu for link.
Committee Members:
Alexandru Bardas, ChairFengjun Li
Bo Luo
Abstract
Activism is a universal concept that has often played a major role in putting an end to injustices and human rights abuses globally. Political activism in specific is a modern day term coined to refer to a form of activism in which a group of people come into collision with a more omnipotent adversary - national or international governments - who often has a purview and control over the very telecommunications infrastructure that is necessary for activists in order to organize and operate. As technology and social media use have become vital to the success of activism movements in the twenty first century, our study focuses on surfacing the technical challenges and the defensive strategies that activists employ during a political revolution. We find that security and privacy behavior and app adoption is influenced by the specific societal and political context in which activists operate. In addition, the impact of a social media blockade or an internet blackout can trigger a series of anti-censorship approaches at scale and cripple activists’ technology use. To a large extent the combination of low tech defensive strategies employed by activists were sufficient against the threats of surveillance, arrests and device confiscation. Throughout our results we surface a number of design principles but also some design tensions that could occur between the security and usability needs of different populations. And thus, we present a set of observations that can help guide technology designers and policy makers.
When & Where:
May 7, 2021 - 9:30 AMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Prasad Kulkarni, ChairDavid Johnson, Co-Chair
Suzanne Shontz
Abstract
Recent advances in machine learning (ML), known as deep neural networks (DNN) or deep learning, have greatly improved the state-of-the-art for many ML tasks, such as image classification (He, Zhang, Ren, & Sun, 2016; Krizhevsky, Sutskever, & Hinton, 2012; LeCun, Bottou, Bengio, & Haffner, 1998; Szegedy et al., 2015; Zeiler & Fergus, 2014), speech recognition (Graves, Mohamed, & Hinton, 2013; Hannun et al., 2014; Hinton et al., 2012), complex games and learning from simple reward signals (Goodfellow et al., 2014; Mnih et al., 2015; Silver et al., 2016), and many other areas as well. NN and ML methods have been applied to the task of autonomously controlling a vehicle with only a camera image input to successfully navigate on road (Bojarski et al., 2016). However, advances in deep learning are not yet applied systematically to this task. In this work I used a simulated environment to implement and compare several methods for controlling autonomous navigation behavior using a standard camera input device to sense environmental state. The simulator contained a simulated car with a camera mounted on the top to gather visual data while being operated by a human controller on a virtual driving environment. The gathered data was used to perform supervised training for building an autonomous controller to drive the same vehicle remotely over a local connection. Reproduced past results that have used simple neural networks and other ML techniques to guide similar test vehicles using a camera. Compared these results with more complex deep neural network controllers, to see if they can improve navigation performance based on past methods on measures of speed, distance, and other performance metrics on unseen simulated road driving tasks.
When & Where:
May 4, 2021 - 2:00 PMZoom Meeting, please contact jgrisafe@ku.edu for link.
Committee Members:
Prasad Kulkarni, ChairDavid Johnson, Co-Chair
Andrew Gill
Abstract
COVID-19 is a disease that spreads from human to human which can be controlled by ensuring proper use of a facial mask. The spread of COVID-19 can be limited if people strictly maintain social distancing and use a facial mask. Very sadly, people are not obeying these rules properly which is speeding the spread of this virus. Detecting the people not obeying the rules and informing the corresponding authorities can be a solution in reducing the spread of Corona virus. The proposed method detects the face from the image correctly and then identifies if it has a mask on it or not. As a surveillance task performer, it can also detect a face along with a mask in motion. It has numerous applications, such as autonomous driving, education, surveillance, and so on.
When & Where:
May 3, 2021 - 1:00 PMZoom Meeting, please contact jgrisafe@ku.edu for link.
Committee Members:
Prasad Kulkarni, ChairDavid Johnson, Co-Chair
Andrew Gill
Abstract
Fatigue and microsleep at the wheel are often the cause of serious accidents and death. Fatigue, in general, is difficult to measure or observe unlike alcohol and drugs, which have clear key indicators and tests that are available easily. Hence, detection of driver’s fatigue and its indication is an active research area. Also, I believe that drowsiness can negatively impact people in working and classroom environments as well. Drowsiness in the workplace especially while working with heavy machinery may result in serious injuries similar to those that occur while driving drowsily. The proposed system for detecting driver drowsiness has a webcam that records the video of the driver and driver’s face is detected in each frame employing image processing techniques. Facial landmarks on the detected face are pointed and subsequently the eye aspect ratio, mouth opening ratio and nose length ratio are computed and depending on their values, drowsiness is detected. If drowsiness is detected, a warning or alarm is sent to the driver from the warning system.
When & Where:
May 3, 2021 - 10:00 AMZoom Meeting, please contact jgrisafe@ku.edu for link.
Committee Members:
Cuncong Zhong, ChairGuanghui Wang
Taejoon Kim
Fengjun Li
Abstract
Fingerspelling in sign language has been the means of communicating technical terms and proper nouns when they do not have dedicated sign language gestures. The automatic recognition of fingerspelling can help resolve communication barriers when interacting with deaf people. The main challenges prevalent in automatic recognition tasks are the ambiguity in the gestures and strong articulation of the hands. The automatic recognition model should address high inter-class visual similarity and high intra-class variation in the gestures. Most of the existing research in fingerspelling recognition has focused on the dataset collected in a controlled environment. The recent collection of a large-scale annotated fingerspelling dataset in the wild, from social media and online platforms, captures the challenges in a real-world scenario. This study focuses on implementing a fine-grained visual attention approach using Transformer models to address the challenges existing in two fingerspelling recognition tasks: multiclass classification of static gestures and sequence-to-sequence prediction of continuous gestures. For a dataset with a single gesture in a controlled environment (multiclass classification), the Transformer decoder employs the textual description of gestures along with image features to achieve fine-grained attention. For the sequence-to-sequence prediction task in the wild dataset, fine-grained attention is attained by utilizing the change in motion of the video frames (optical flow) in sequential context-based attention along with a Transformer encoder model. The unsegmented continuous video dataset is jointly trained by balancing the Connectionist Temporal Classification (CTC) loss and maximum-entropy loss. The proposed methodologies outperform state-of-the-art performance in both datasets. In comparison to the previous work for static gestures in fingerspelling recognition, the proposed approach employs multimodal fine-grained visual categorization. The state-of-the-art model in sequence-to-sequence prediction employs an iterative zooming mechanism for fine-grained attention whereas the proposed method is able to capture better fine-grained attention in a single iteration.
When & Where:
April 30, 2021 - 1:00 PMZoom Meeting, please contact jgrisafe@ku.edu for link.
Committee Members:
Morteza Hashemi, ChairDavid Johnson
Taejoon Kim
Abstract
With the rapid development of machine learning (ML) and deep learning (DL) methodologies, DL methods can be leveraged for wireless network reconfigurability and channel modeling. While deep learning-based methods have been applied in a few wireless network use cases, there is still much to be explored. In this project, we focus on the application of deep learning methods for two scenarios. In the first scenario, a user transmitter was moving randomly within a campus area, and at certain spots sending wireless signals that were received by multiple antennas. We construct an active deep learning architecture to predict user locations from received signals after dimensionality reduction, and analyze 4 traditional query strategies for active learning to improve the efficiency of utilizing labeled data. We propose a new location-based query strategy that considers both spatial density and model uncertainty when selecting samples to label. We show that the proposed query strategy outperforms all the existing strategies. In the second scenario, a reconfigurable intelligent surface (RIS) containing 4096 tunable cells reflects signals from a transmitter to users in an office for better performance. We use the training data of one user's received signals under different RIS configurations to learn the impact behavior of the RIS on the wireless channel. Based on the context and experience from the first scenario, we build a DL neural network that maps RIS configurations to received signal estimations. In the second phase, the loss function was customized towards our final evaluation formula to obtain the optimum configuration array for a user. We propose and build a customized DL pipeline that automatically learns the behavior of RIS on received signals, and generates the optimal RIS configuration array for each of the 50 test users.
PAST DEFENSE NOTICES
When & Where:
March 8, 2021 - 10:30 AMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Alex Bardas, ChairFengjun Li
Bo Luo
Abstract
Android's fragmented ecosystem makes the delivery of security updates and OS upgrades cumbersome and complex. While Google initiated various projects such as Android One, Project Treble, and Project Mainline to address this problem, and other involved entities (e.g., chipset vendors, manufacturers, carriers) continuously strive to improve their processes, it is still unclear how effective these efforts are on the delivery of updates to supported end-user devices. In this paper, we perform an extensive quantitative study (August 2015 to December 2019) to measure the Android security updates and OS upgrades rollout process. Our study leverages multiple data sources: the Android Open Source Project (AOSP), device manufacturers, and the top four U.S. carriers (AT\&T, Verizon, T-Mobile, and Sprint). Furthermore, we analyze an end-user dataset captured in 2019 (152M anonymized HTTP requests associated with 9.1M unique user identifiers) from a U.S.-based social network. Our findings include unique measurements that, due to the fragmented and inconsistent ecosystem, were previously challenging to perform. For example, manufacturers and carriers introduce a median latency of 24 days before rolling out security updates, with an additional median delay of 11 days before end devices update. We show that these values alter per carrier-manufacturer relationship, yet do not alter greatly based on a model's age. Our results also delve into the effectiveness of current Android projects. For instance, security updates for Treble devices are available on average 7 days faster than for non-Treble devices. While this constitutes an improvement, the security update delay for Treble devices still averages 19 days.
When & Where:
February 8, 2021 - 10:00 AMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Prasad Kulkarni, ChairMorteza Hashemi
Taejoon Kim
Alessandro Salandrino
Elaina Sutley
Abstract
When & Where:
January 29, 2021 - 10:30 AMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Taejoon Kim, ChairBo Luo
Fengjun Li
Cuncong Zhong
Guanghui Wang
Abstract
Crowd estimation is an integral part of crowd analysis. It plays an important role in event management of huge gatherings like Hajj, sporting, and musical events or political rallies. Automated crowd count can lead to better and effective management of such events and prevent any unwanted incident. Crowd estimation is an active research problem due to different challenges pertaining to large perspective, huge variance in scale and image resolution, severe occlusions and dense crowd-like cluttered background regions. Current approaches cannot handle huge crowd diversity well and thus perform poorly in cases ranging from extreme low to high crowd-density, thus, leading to crowd underestimation or overestimation. Also, manual crowd counting subjects to very slow and inaccurate results due to the complex issues as mentioned above. To address the major issues and challenges in the crowd counting domain, we separately investigate two different types of input data: uni-modal (Image) and multi-modal (Image and Audio).
In the uni-modal setting, we propose and analyze four novel end-to-end crowd counting networks, ranging from multi-scale fusion-based models to uniscale one-pass and two-pass multi-task models. The multi-scale networks also employ the attention mechanism to enhance the model efficacy. On the other hand, the uni-scale models are equipped with novel and simple-yet-effective patch re-scaling module (PRM) that functions identical but lightweight in comparison to the multi-scale approaches. Experimental evaluation demonstrates that the proposed networks outperform the state-of-the-art methods in majority cases on four different benchmark datasets with up to 12.6% improvement in terms of the RMSE evaluation metric. Better cross-dataset performance also validates the better generalization ability of our schemes. For the multimodal input, the effective feature-extraction (FE) and strong information fusion between two modalities remain a big challenge. Thus, the aim in the multimodal environment is to investigate different fusion techniques with improved FE mechanism for better crowd estimation. The multi-scale uni-modal attention networks are also proven to be more effective in other deep leaning domains, as applied successfully on seven different scene-text recognition datasets with better performance.
When & Where:
January 29, 2021 - 2:00 AMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Fengjun Li, ChairAlex Bardas
Bo Luo
Cuncong Zhong
Mei Liu
Abstract
Machine learning (ML) is transforming a wide range of applications, promising to bring immense economic and social benefits. However, it also raises substantial security and privacy challenges. In this dissertation we describe a framework for efficient, collaborative and secure ML training using a federation of client devices that jointly train a ML model using their private datasets in a process called Federated Learning (FL). First, we present the design of a blockchain-enabled Privacy-preserving Federated Transfer Learning (PPFTL) framework for resource-constrained IoT applications. PPFTL addresses the privacy challenges of FL and improves efficiency and effectiveness through model personalization. The framework overcomes the computational limitation of on-device training and the communication cost of transmitting high-dimensional data or feature vectors to a server for training. Instead, the resource-constrained devices jointly learn a global model by sharing their local model updates. To prevent information leakage about the privately-held data from the shared model parameters, the individual client updates are homomorphically encrypted and aggregated in a privacy-preserving manner so that the server only learns the aggregated update to refine the global model. The blockchain provides provenance of the model updates during the training process, makes contribution-based incentive mechanisms deployable, and supports traceability, accountability and verification of the transactions so that malformed or malicious updates can be identified and traced to the offending source. The framework implements model personalization approaches (e.g. fine-tuning) to adapt the global model more closely to the individual client's data distribution.
In the second part of the dissertation, we turn our attention to the limitations of existing FL algorithms in the presence of adversarial clients who may carry out poisoning attacks against the FL model. We propose a privacy-preserving defense, named CONTRA, to mitigate data poisoning attacks and provide a guaranteed level of accuracy under attack. The defense strategy identifies malicious participants based on the cosine similarity of their encrypted gradient contributions and removes them from FL training. We report the effectiveness of the proposed scheme for IID and non-IID data distributions. To protect data privacy, the clients' updates are combined using secure multi-party computation (MPC)-based aggregation so that the server only learns the aggregated model update without violating the privacy of users' contributions.
When & Where:
January 28, 2021 - 1:00 PMZoom Meeting, please contact jgrisafe@ku.edu for link
Committee Members:
Prasad Kulkarni, ChairShawn Keshmiri, (Co-Chair)
Alex Bardas
Morteza Hashemi
Abstract
Swarms of unmanned aerial systems (UASs) usage is becoming more prevalent in the world. Many private companies and government agencies are actively developing analytical and technological solutions for multi-agent cooperative swarm of UASs. However, majority of existing research focuses on developing guidance, navigation, and control (GNC) algorithms for swarm of UASs and proof of stability and robustness of those algorithms. In addition to profound challenges in control of swarm of UASs, a reliable and fast intercommunication between UASs is one of the vital conditions for success of any swarm. Many modern UASs have high inertia and fly at high speeds which means if latency or throughput are too low in swarms, there is a higher risk for catastrophic failure due to intercollision within the swarm. This work presents solutions for scaling number of collaborative agents in swarm of UASs using frequency-based hierarchy. This work identifies shortcomings and discusses traditional swarm communication systems and how they rely on a single frequency that will handle distribution of information to all or some parts of a swarm. These systems typically use an ad-hoc network to transfer data locally, on the single frequency, between agents without the need of existing communication infrastructure. While this does allow agents the flexibility of movement without concern for disconnecting from the network and managing only neighboring communications, it doesn’t necessarily scale to larger swarms. In those large swarms, for example, information from the outer agents will be routed to the inner agents. This will cause inner agents, critical to the stability of a swarm, to spend more time routing information than transmitting their state information. This will lead to instability as the inner agents’ states are not known to the rest of the swarm. Even if an ad-hoc network is not used (e.g. an Everyone-to-Everyone network), the frequency itself has an upper limit to the amount of data that it can send reliably before bandwidth constraints or general interference causes information to arrive too late or not at all.