Year: 2023
Type: dissertation
Abstract: As machine learning and deep learning models become ubiquitous, it is inevitable that there will be attempts to exploit such models in various attack scenarios. For example, in a steganographic based attack, information would be hidden in a learning model, which might then be used to gain unauthorized access to a computer, or for other malicious purposes. In this research, we determine the steganographic capacity of various classic machine learning and deep learning models. Specifically, we determine the number of low-order bits of the trained parameters of a given model that can be altered without significantly affecting the performance of the model. We find that the steganographic capacity of learning models is surprisingly high, and that there tends to be a clear threshold after which model performance rapidly degrades.
Author: Lei Zhang
Institution: San Jose State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4380786307
