Transportability of Adversarial Attack

The aim of this thesis is to leverage machine learning algorithms to introduce adversarial cloaks, or noise, into images containing real individuals and assess whether the resulting images still accurately represent human subjects. Specifically, I seek to discern whether images featuring avatars, placed in the same spatial configuration as humans and subjected to similar perturbations, can be effectively used to generate adversarial examples capable of fully masking a person. By analyzing the probability of correctly identifying personhood, I strive to establish a method of utilizing virtual avatars to provide increased privacy for users. This research has significant implications for advancing privacy and security measures pertaining to personal image data. Through the exploration of machine learning techniques and the application of adversarial perturbations/cloaks, this study aims to contribute to the development of robust methods for human recognition and differentiation in visual data. The findings of this research hold the potential to impact diverse fields, including computer vision, privacy protection, and human detection technologies