Workshops Archive

A Threat Model for Sensitive-Data Workloads on Multi-Tenant AI Supercomputers


Workshop: 5th International Workshop on Cyber Security in High Performance Computing

Authors: Sadaf R. Alam (University of Bristol, Bristol Centre for Supercomputing); Martin O'Reilly, Matt Craddock, and Jim Madge (The Alan Turing Institute); Matt Penn (Kings College London); George Svarovsky (University College London); Jake Watson (University of Bristol); and Babatunde Oyewo (University of Cambridge)

Abstract: Sensitive-data research increasingly requires the GPU scale of national AI supercomputers, yet the governed multi-tenant handling routine in cloud environments has not been established in supercomputing. Extending a Trusted Research Environment (TRE), or Secure Data Environment (SDE), onto such a system distributes security responsibility across three parties---a capacity owner, an infrastructure provider, and a TRE/SDE operator that governs data on infrastructure it neither owns nor operates. We present a threat model for this setting, expressed as attack trees and grounded in FRIDGE, a satellite-TRE architecture deployed as multi-tenant Kubernetes on two bare-metal national systems. It covers three adversaries---a privileged host operator, a malicious co-tenant, and a rogue researcher---and the supercomputing-specific attack surfaces conventional TRE work omits, mapping each attack path to a mitigation and its owning party. Underpinning the design is portable governance across independently-operated TRE/SDE platforms. This is work in progress.


Back to 5th International Workshop on Cyber Security in High Performance Computing Archive Listing

Back to Full Workshop Archive Listing