Available at: https://digitalcommons.calpoly.edu/theses/3207
Date of Award
12-2025
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Alex Dekhtyar
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Generative AI is powerful yet controversial technology capable of spreading misinformation, non-consensual imagery, and offensive materials when not properly safeguarded against such abuse. Synthetic image alteration and generation have always been abused, from photo editing software to Generative Adversarial Nets, and now Diffusion Models. The Internet Watch Foundation and the National Center for Missing and Endangered Children have reported a consistent increase in reported AI generated child sexual assault materials (CSAM) and sextortion scams since the introduction of DMs. New digital forensics tools must evolve to adapt to the ever-changing landscape of AIG CSAM abuse. This work provides a comprehensive evaluation of signature-based model attribution for finetuned versions of Stable Diffusion and LoRAs, providing insight into detectability of models under a variety of circumstances and the effectiveness of the attribution methodology. Model signatures are created from and evaluated on a benign dataset of images from various finetuned Stable Diffusion models and associated LoRAs. Unseen image residuals are compared to model fingerprints through the use of comparison windows, mxm sections of an image or fingerprint, and a distance metric is calculated to attribute a given image to the model fingerprint it is most similar to. This thesis provides insight into the most effective size and location of a comparison window and compares performance across different signature creation methods in binary and multiclass classification. The results of this thesis indicate that signature-based attribution can be applied at a higher level of granularity than previously explored.