Advancing the Efficient and Trustworthy AI on the Edge

Abstract

The proliferation of mobile devices, IoT sensors, and wearable technologies has catalyzed a shift in artificial intelligence (AI) from centralized cloud computing to distributed intelligence at the edge. Edge AI enables models to be trained and deployed closer to data sources, thereby improving responsiveness and preserving user privacy. However, this paradigm introduces two fundamental challenges: efficiency—how to train and deploy AI models within the constrained resources of edge devices—and trustworthiness—how to ensure the reliability, robustness, and privacy of AI systems in real-world deployments.This dissertation aims to advance the frontiers of efficient and trustworthy edge AI by developing algorithmic and system-level innovations across three critical dimensions: collaborative learning, on-device model adaptation, and trustworthy edge AI.First, I address the challenges of heterogeneity and inefficiency in federated learning (FL), a core paradigm for privacy-preserving, collaborative edge AI. I propose FedSEA, a semi-asynchronous FL framework that accommodates straggling and diverse clients through dynamic scheduling and heterogeneity-aware aggregation. To extend FL’s applicability to cross-silo and multi-modal scenarios, I also introduce a one-shot and few-shot vertical FL approach that reduces communication overhead while enabling learning with limited sample overlap.Second, I explore on-device adaptation of foundation models such as large language models (LLMs), which are traditionally too resource-intensive for edge deployment. I develop FedBPT, a gradient-free, black-box prompt tuning framework that supports personalized model adaptation using only inference-level feedback. Furthermore, I propose a knowledge graph tuning method that incorporates user-defined knowledge structures for real-time, human-in-the-loop personalization. These techniques allow for efficient and privacy-respecting adaptation of large models under edge constraints.Third, I focus on enhancing the trustworthiness of collaborative edge intelligence. I introduce Soteria, a representation-level privacy defense that prevents information leakage while preserving model utility. I also present FL-WBC, a client-side robust training strategy that defends against model poisoning attacks by reorganizing local gradients. Both methods offer theoretical guarantees and empirical effectiveness under adversarial conditions.Together, these contributions establish a comprehensive framework for building edge AI systems that are not only efficient and scalable but also privacy-preserving and robust. This work lays the foundation for deploying intelligent applications that are aligned with user needs and constraints in real-world edge environments.