Neural network security and optimization for single-person authentication using electroencephalogram data

Andre, Naomi, author
Simske, Steve, advisor
Mueller, Jennifer, committee member
Lyons, Michael, committee member
Journal Title
Journal ISSN
Volume Title
Security is an important focus for devices that use biometric data, and as such security around authentication needs to be considered. This is true for brain-computer interfaces (BCIs), which often use electroencephalogram (EEG) data as inputs and neural network classification to determine their function. EEG data can also serve as a form of biometric authentication, which would contribute to the security of these devices. Neural networks have also used a method known as ablation to improve their efficiency. In light of this info, the goal of this research is to determine whether neural network ablation can also be used as a method to improve security by reducing a network's learning capabilities to include authenticating only a given target, and preventing adversaries from training new data to be authenticated. Data on the change in entropy of weight values of the networks after training was also collected for the purpose of determining patterns in weight distribution. Results from a set of ablated networks to a set of baseline (non-ablated) networks for five targets chosen randomly from a data set of 12 people were compared. The results found that ablated maintained accuracy through the ablation process, but that they did not perform as well as the baseline networks. Change in performance between single-target authentication and target-plus-invader authentication was also examined, but no significant results were found. Furthermore, the change in entropy differed between both baseline networks and ablated networks, as well as between single-target authentication and target-plus-invader authentication for all networks. Ablation was determined to have potential for security applications that need to be expanded on, and weight distribution was found to have some correlation with the complexity of an input to a network.
Includes bibliographical references.
2022 Fall.
Rights Access
Associated Publications