Deep Ensemble Learning for Robust Defense Against Adversaries
Open Access
- Author:
- Adams, Ethan
- Area of Honors:
- Information Sciences and Technology
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Mahdi Nasereddin, Thesis Supervisor
Sandy Feinstein, Thesis Honors Advisor - Keywords:
- AI
Artificial Intelligence
ML
Machine Learning
TensorFlow
Keras
Python
AI Security
Ensemble Learning
Deep Learning
Neural Networks
Adversarial
Adversaries - Abstract:
- Several measures have been taken to increase the security of AI models against adversarial inputs in response to growing reliance on the technology. While many data-driven solutions have been created, they do little to address the structural problems of the model. A novel model is tested with an ensemble structure, which subsamples its dataset, assigning portions to each sub-model in a nested configuration. The model is then tested against a control model trained on the overall dataset. The ensemble solution showed significant improvements in accuracy across multiple testing procedures when compared to the control model. Ensemble structuring is shown to be an effective method for improving model robustness without extending training time or processing workload required.
Accessible Version in Progress
We're generating an accessible version of this file to meet ADA Title II requirements. This process may take up to one hour. Please return later to access the accessible copy once it's ready.
You can still download the current version by clicking "OK".
What's happening:
An accessible PDF is being generated using Adobe with AI used to generate alternative text (alt text) for images in the PDF.