Can we trust generative AI?

• Main Author: Mu, Zhan
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering; Computer and information science
• Online Access: https://hdl.handle.net/10356/182919
• Institution: Nanyang Technological University

Generative Artificial Intelligence (GAI) models have demonstrated remarkable capabilities across various domains, yet their robustness remains a significant challenge, especially when exposed to adversarial attacks. This project aims to address the robustness issues in GAI models by evaluating their performance under adversarial conditions. Specifically, we utilize the COCO dataset to gen- erate adversarial examples using a variety of attack methods, including PGD, FMA, MULTIMODAL-FUSION, and DI-PGD. BLIP-large serves as the pre- trained model to generate adversarial datasets, which are subsequently used to attack other GAI models, including UniDiffuser, LLaVA, BLIP2-OPT, MiniGPT- 4, and BLIP-base. To assess the impact of adversarial attacks, we adopt eval- uation metrics such as CLIP score and textual cosine similarity to measure the differences between the captions generated on adversarial datasets and the orig- inal captions as well as their alignment with the original image content. This project not only reveals the vulnerabilities of current GAI models under ad- versarial scenarios but also establishes a benchmark for adversarial attacks on the latest GAI models, providing new references for research in this field. The findings of this study lay a solid foundation for future efforts to enhance the robustness of GAI models.