Towards a more trustworthy generative artificial intelligence

• 主要作者: Cheong, Ben Wee Joon
• 其他作者: Alex Chichung Kot
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2024
• 主題: Computer and Information Science; GAN; Hallucinations
• 在線閱讀: https://hdl.handle.net/10356/181657
• 機構: Nanyang Technological University

In recent years, the rapid increase of available Generative Artificial Intelligence (AI) models has revolutionized various domains. From AI image generation to intelligent natural language reasoning like Open AI’s ChatGPT and Google Gemini. These models fuelled by advancements in deep learning have demonstrated unprecedented capabilities in generating life-like content. However, with these advancements comes new challenges, especially in ensuring the robustness and reliability of such models when faced with adversarial attacks. Adversarial attacks are attacks that exploits the vulnerabilities of Generative AI models by subtly altering the input data, also known as perturbations, to deceive the models into making hallucinated, distorted, or incorrect predictions. Vision Language Models (VLM) are an example of Generative AI that leverages complex neural architectures, combining the capabilities of both Computer Vision (CV) and Natural Language Processing (NLP) modalities to generate captions or descriptions of images or to produce corresponding visual content from textual prompts. In the context of VLM, adversarial attacks can come in the form of perturbated image inputs ranging from subtle to severe using techniques such as gaussian blur, dithering, and contrast manipulation leading to hallucinated image captioning or description. Hallucinations refer to erroneous or implausible outputs generated by AI models, particularly in generative tasks. In this context, hallucinations manifest as generated images or captions that deviate significantly from the intended input. Motivated by the need to address this vulnerability, this project systematically evaluates VLMs’ robustness under a variety of adversarial perturbations, with a focus on benchmarking hallucinations. Existing hallucination benchmark work faces challenges, such as limited task specificity and lack of depth in assessing VLM responses across diverse perturbations. By addressing these gaps, this study aims to enhance our understanding of VLM limitations and contribute to the development of more robust and reliable vision-language models, setting a foundation for improving AI models’ resilience to real-world distortions.