Trustworthy AI in the Modern Era: Theories and Applications

Abstract

Artificial intelligence (AI) has become increasingly prevalent in various domains, thereby highlighting the importance of understanding and ensuring its safety. This work focuses on enhancing AI trustworthiness, delving into theoretical foundations and practical algorithms. Particularly, I focus on four interconnected critical components of modern AI: model security, fairness, explainability, and data privacy. For model security, I aim to ensure that the model integrity and behavior are not compromised against potential malicious attacks, especially model backdoor and stealing attacks. I propose a unified framework named "model privacy'' to analyze those attacks, leading to a fundamental understanding and inspiring better design of defense mechanisms. For fairness, I study the group fairness of the learned model in a decentralized setting, ensuring the benefits of AI technologies are equitably enjoyed by everyone regardless of their gender, race, and other diverse backgrounds. For model explainability, I address the problem of how much we can prune an AI model without sacrificing accuracy. By leveraging a sparsity index based on the ℓ𝑞-norm of model parameters, I quantify the compressibility of a model through its inherent sparsity. Furthermore, an adaptive iterative pruning algorithm is proposed and achieves the state-of-the-art performance. Lastly, for data privacy, I strive to protect confidential individual information from being revealed. To achieve this goal, I propose a private data collection mechanism named "subset privacy'', which reports a set containing the truth. With subset privacy, the exact value of truth is inaccessible to others while the data analyst can still effectively extract useful information from the privatized data.