Security and Privacy of Locally-Run AI Models (SPLAIM)

The research project SPLAIM (Security and Privacy of Locally-Run AI Models) enables Swedish organizations to deploy large language models (LLMs) securely, privately, and efficiently on their own infrastructure. Today, many enterprises cannot fully leverage cloud-based AI services because of strict requirements for data protection, compliance, and digital sovereignty. SPLAIM addresses this structural barrier by developing the scientific foundations for integrating security, privacy, governance, and resource efficiency in local AI systems that operate under real organizational and regulatory constraints.

The project combines three intertwined research strands: (1) systematic analysis of security and privacy risks for locally deployed LLMs; (2) design and evaluation of privacy-preserving fine-tuning and retrieval-augmented generation; and (3) empirical studies and co-design of adoption and governance models that embed organizational, regulatory, and usability requirements into system design. 

SPLAIM follows an iterative co-design cycle with requirement engineering, technical development, prototyping, and evaluation across realistic use cases. The project aims to deliver validated threat models, methods, and tools for privacy-preserving local LLM pipelines, along with benchmarks and guidelines for secure and resource-efficient deployment.