Trustworthy and Reliability in Centralized and Distributed AI

AI is becoming more and more popular, with the ability to learn from billions of data points and collect users' data to create a personalized experience for every user. This leads to the critical issues of privacy and security in centralized and distributed AI. In centralized AI, a single model is trained on aggregated data in one location, which can lead to privacy risks because sensitive information may be memorized and embedded in the model. In distributed AI, multiple clients collaboratively train a shared model without sharing raw data, but the system must handle security risks, such as malicious clients that can corrupt model updates. This thesis proposes two mechanisms to enhance the reliability and trustworthiness of both centralized and distributed AI. We first address safety in centralized Large Reasoning Models (LRMs). While the Chain-of-Thought (CoT) of LRMs provides unprecedented logical depth, they also create significant vulnerabilities by memorizing and potentially leaking sensitive or copyrighted training data through intermediate reasoning steps. We introduce FRUL, a novel Feature Replacement-Aware Unlearning framework designed to sanitize centralized models' post-deployment. By leveraging multiple LLMs, Retrieval-Augmented Generation (RAG), and FRUL unlearning loss, we demonstrate that centralized models can effectively forget insecure or private data while preserving their core reasoning integrity. In the second phase, we transition to robustness within distributed AI. We address the challenge of securing AI models from malicious users who attempt to corrupt the shared models through poisoning attacks. We propose BR-MTRL, a Byzantine-resilient multi-task representation learning framework. By decoupling the architecture into shared global representations and local personalized heads, and implementing robust aggregation methods such as Geometric Median and Krum, we ensure the system remains resilient against adversarial clients. Overall, this research provides a comprehensive blueprint for Reliable and Trustworthy AI.

Committee: Mengdi Huai (major professor), Shana Moothedath (major professor) and Qi Li