Advancing adversarial and LLM robustness in trustworthy AI: a comprehensive survey

Despite significant advancements in various fields in recent years, artificial intelligence (AI) models have demonstrated strong performance and broad application potential. However, they still face numerous challenges in terms of security and robustness in practical applications. Among these, robustness stands out as a critical factor contributing to the perceived untrustworthiness of AI models and remains a major barrier to their widespread adoption. Moreover, most current AI models are designed with a black-box structure, lacking sufficient explainability, which makes it difficult for researchers to understand their decision-making mechanisms. This ‘invisible’ nature not only limits the ability to predict model behavior but also increases the instability of models in complex and unknown environments. In this paper, we systematically review the evaluation methods and enhancement strategies for AI model robustness from multiple perspectives: (1) We identify the primary issues and technical challenges in the robustness of current AI models. (2) We explore the connections and distinctions between core concepts of trustworthy AI. (3) We summarize the development of robustness evaluation in recent years from both the perspective of robustness evaluation metrics and methods. (4) We examine robustness enhancement methods across different stages of the AI model lifecycle: data preprocessing, training, model architecture design, and post-processing. (5) We focus on hallucinations and other robustness issues faced by generative large language models (LLMs), summarizing current research progress and mitigation strategies. (6) Finally, we discuss open questions and future research directions in the field of AI model robustness.