Robust Defenses for Artistic and Personal Data Against Unauthorized AI Exploitation
Date Issued
August 2026
Author(s)
Meerza, Syed Irfan Ali
The University of Tennessee
Advisor(s)
Jian Liu
Additional Advisor(s)
Jinyuan Sun
Feiyi Wang
FNU Suya
Abstract
With the rapid advancement of Artificial Intelligence (AI) in image, audio, and multimedia generation, growing concerns have emerged regarding the unauthorized use of personal data and creative works. Modern AI systems can be trained on publicly available content, imitate artistic styles, edit existing media, and generate new content that closely resembles the original creators' work. These capabilities raise important questions about ownership, consent, and creative control. This dissertation investigates AI-resistant protection techniques that enable individuals and creators to safeguard their digital assets, particularly images and music, from unauthorized learning, manipulation, and imitation by AI systems.
Specifically, this dissertation addresses four fundamental challenges: (1) how to modify images so that AI models cannot effectively learn from them while preserving their visual quality; (2) how to prevent generative music models from learning and reproducing copyrighted musical content; (3) how to protect music from AI-driven editing, remixing, and style imitation in addition to unauthorized training; and (4) how to defend visual artwork against style mimicry by modern image-generation systems.
First, DiffUE is introduced as a framework that protects images by introducing carefully designed modifications in a representation space that captures their semantic content. These modifications preserve the visual appearance of the image while preventing AI models from learning useful information from it, even when adversaries employ retraining strategies or attempt to remove the protection.
Second, HarmonyCloak embeds imperceptible perturbations into music recordings to prevent generative AI models from learning meaningful representations of protected compositions. As a result, models trained on protected music are significantly less capable of reproducing the original musical content, while the listening experience remains unchanged for human audiences.
Third, MusicShield extends protection beyond training-time defense to address AI-assisted editing, remixing, and style imitation. By strategically altering the representations learned by generative models, MusicShield reduces their ability to manipulate protected tracks or generate music that closely resembles the original artist's style.
Finally, Mirage is presented as a robust artwork-protection framework that introduces imperceptible and resilient modifications into style-sensitive image representations. These modifications remain effective even when adversaries apply advanced image restoration techniques and significantly reduce the ability of generative AI systems to imitate an artist's unique visual style.
Together, these contributions advance the state of AI-resistant media protection and provide practical tools that help creators maintain ownership, autonomy, and creative integrity in an increasingly AI-driven world.
Degree
Doctor of Philosophy
Major
Computer Engineering
Embargo Date
August 1, 2027