Modeling and Evaluation of Continuous Authentication with Keystroke Dynamics

Date

2023-7-17

Author

Bıyık, Mustafa

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

283
views

255
downloads

Cite This

Authentication with single- and multi-factor credentials alone cannot prevent sophisticated identity theft attacks like account takeovers and session hijackings. Companies and individuals may suffer catastrophic losses and critical infrastructure may be destroyed as a result of these attacks. Detecting identity theft and session hijacking attacks with continuous authentication schemes is emerging as a promising technique for extending traditional identity and access management (IAM) strategies. As a first step, user authentication and authorization are key components of IAM strategies that protect data security and privacy. A continuous authentication system verifies the identity of a user repeatedly or continuously based on behavior patterns unique to each user rather than simply identifying users through their credentials one time, as is the case in classical authentication. As a result, due to its hard-to-imitate nature, personal keystroke dynamics can be used for identity authentication. Also, verification can be done without affecting input, and authentication is made in the background without affecting users' input. Two different methods are used in this thesis. The first approach is that the behavioral characteristics of the user for keystroke dynamics are modeled as a Gamma distribution due to the right skewness of keystroke data distributions. For each user in the system, a reference Gamma model and observed Gamma models are generated from keystroke dynamics features such as the time between two key presses. The error distances measured with Jensen-Shannon (JS) distance metric between the reference models and the observed model are used for user identification. M-ary hypothesis testing method is used in the decision stage of user detection. Also, simulations and different scenarios are done to evaluate the performance of the proposed model in the thesis. The second approach is based on the conversion of keystroke data with a constant length into a 2D image. By using this transformation process, we can train images based on convolutional neural networks (CNNs) while maintaining all the characteristics of the behavioral signal. Popular CNN architectures which are Resnet-18, DenseNet-121, EffifcientNetB0, ShuffleNet, and a Basic CNN model created for the research are used. Implementation of Gamma model distribution is easier. On the other side, CNN provides higher accuracy. In this thesis, two different authentication methods are evaluated in terms of accuracy and efficiency on different datasets.

Subject Keywords

continuous authentication, keystroke dynamics, Gamma distribution, hypothesis testing, CNN

URI

https://hdl.handle.net/11511/104837

Collections

Graduate School of Natural and Applied Sciences, Thesis