Show simple item record

Adversarially Robust Machine Learning Approaches for Edge-based Applications

dc.contributor.advisorGokhale, Aniruddha
dc.creatorCanady, Robert Edward
dc.date.accessioned2024-08-15T19:00:40Z
dc.date.created2024-08
dc.date.issued2024-07-12
dc.date.submittedAugust 2024
dc.identifier.urihttp://hdl.handle.net/1803/19217
dc.description.abstractLeveraging deep learning models at the edge is becoming increasingly more prevalent due to their performance as well as security, privacy, or environmental concerns. These models can be used for applications ranging from surveillance, autonomous vehicles, personal assistants, etc. Despite their success, utilizing these edge-based models comes with several challenges. First, the edge is characterized by having constraints on compute ability, storage, and power. Thus a smaller model is needed, which is typically done by either pruning the model down in size or quantizing the model's weights and biases. This needs to be done while minimizing the loss of accuracy. Second, the edge is also characterized by having heterogeneous resources. Deep learning models executing on the heterogeneous edge will have different accuracy, execution time, and energy consumption depending on the device. This makes it difficult to decide on the most optimal combination of device and model to use. Third, deep learning models are vulnerable to adversarially crafted perturbations that cause the model to be ineffective by leading it to incorrect inferences. The adversarial vulnerability of edge-based deep learning hinders their deployment, especially in safety-critical scenarios. Defending against these attacks necessitates the use of computationally expensive approaches that cannot be utilized on the edge. There has been much research into each individual area, but the intersection or combination of them still has many open problems. To address these problems, this proposal investigates ways to integrate several different edge and cloud-based deep learning approaches and other tools such as adversarial training, sensor fusion, knowledge distillation, Apache TVM, etc. Our goal is to develop novel techniques that ease the design and deployment of robust edge-based deep learning applications. The approach relies on profiling of the application using simulated and deployed statistics on model robustness and resource consumption, which is then adaptively refined based on the model's performance until the algorithm converges.
dc.format.mimetypeapplication/pdf
dc.language.isoen
dc.subjectAdversarial Machine Learning, Edge Computing
dc.titleAdversarially Robust Machine Learning Approaches for Edge-based Applications
dc.typeThesis
dc.date.updated2024-08-15T19:00:40Z
dc.type.materialtext
thesis.degree.namePhD
thesis.degree.levelDoctoral
thesis.degree.disciplineElectrical and Computer Engineering
thesis.degree.grantorVanderbilt University Graduate School
local.embargo.terms2025-08-01
local.embargo.lift2025-08-01
dc.creator.orcid0009-0007-3638-5116
dc.contributor.committeeChairGokhale, Aniruddha


Files in this item

Icon

This item appears in the following Collection(s)

Show simple item record