Input Structure Based Optimization for Privacy Preserving AI Systems

As Artificial Intelligence (AI) systems become increasingly integrated into critical domains, ensuring privacy-preserving model design and system deployment has become a pressing priority. Safeguarding both sensitive user data and proprietary model parameters is critical throughout the AI model and system, from data acquisition and pre-processing to model inference and deployment. However, existing privacy-preserving frameworks face several limitations, including fragmented data ownership, incomplete protection across system stages, substantial computational overhead, and poor scalability to modern architectures such as large language models. This dissertation explores a unifying optimization strategy centered on input structure design to address these challenges. The core idea is to systematically exploit the structural properties of input data, such as sparsity, alignment, batching, and modularity, to minimize redundant cryptographic operations and reduce the computational and communication costs of privacy-preserving machine learning. By adapting protocol designs to the input structure, the proposed methods improve the efficiency and scalability of secure inference and multiparty collaboration. Together, these contributions establish input structure optimization as a critical tool for building practical and robust privacy-preserving AI systems.