This PhD Thesis plans to leverage server-side confidential computing devices to secure the processes of training and inference in decentralized machine learning. Operating on distributed data, be it for inference or for training a distributed model holds the promise for more private forms of machine learning. Instead of having a single or a few service providers that collect user data, data can remain on user devices without being collected in large data silos. In spite of this, decentralized learning does not automatically guarantee privacy. Indeed, in a distributed setting any participant becomes a potential attacker.
In this setting, the integration of confidential computing into artificial intelligence (AI) systems is becoming increasingly important [1] as organizations adopt AI for handling sensitive data in sectors like healthcare, finance, and defense. Retrieval-Augmented Generation (RAG) systems, which combine information retrieval with generative models, such as large language models (LLMs), are a popular solution for tasks like document summarization, chatbot development, and recommendation engines. However, they are often deployed in cloud environments, which introduces security risks, including data leakage, intellectual property (IP) theft, and model tampering.
The brand new Intel TDX CPU or AMD SEV-SNP features provide a promising solution by securing RAG systems through the isolation of both AI models and data within trusted execution environments (TEEs), even when deployed in multi-cloud or untrusted cloud infrastructures. Confidential computing features like TDX or SEV-SNP protect sensitive workloads by encrypting data during processing (data-in-use), while also ensuring model privacy and integrity using cryptographic attestation [2,3]
The student will leverage the Wide team's expertise in Operating Systems and Hypervisors, on trusted execution environments [4], and decentralized machine learning [5,6] to design a Trusted Virtual Machine Monitor (VMM) that will manage and secure Intel TDX-enabled Retrieval-Augmented Generation (RAG) systems, addressing key research objectives. Designing a Trusted VMM presents significant challenges due to the stringent requirements for security, performance, and scalability.
A first challenge lies in ensuring that the VMM operates transparently with the guest operating systems (OS). This entails providing essential security services—such as isolation and resource management—without modifying the guest OS or requiring it to be aware of the underlying TDX-aware VMM. Ensuring this transparency is critical for compatibility with existing systems and workloads, as modifying the guest OS would introduce substantial overhead and deployment complexity.
A second challenge results from the fact that not all available hardware will be equipped with confidential-computing devices. This results from the presence of non-trusted or incompatible server-side devices. To this end, we plan to leverage clustering approaches that can make it possible to leverage groups of devices that can communicate with each other. However, naive clustering would make the system very vulnerable to attacks. As a result, we plan to leverage a hybrid solution in which the data managed by a model is split between a protected and an unprotected part, leading to an inference or training process that is split between trusted and non trusted devices.
