Side and covert channels (referred to collectively as illicit channels) are an insidious affliction of high security systems brought about by the unwanted and unregulated sharing of state amongst processes.

Illicit channels can be effectively broken through isolation, which limits the degree by which processes can interact. The drawback of using isolation as a general mitigation against illicit channels is that it can be very wasteful when employed naively. In particular, permanently isolating every tenant of a public cloud service to its own separate machine would completely undermine the economics of cloud computing, as it would remove the advantages of consolidation.

On closer inspection, it transpires that only a subset of a tenant's activities are sufficiently security sensitive to merit strong isolation. Moreover, it is not generally necessary to maintain isolation indefinitely, nor is it given that isolation must always be procured at the machine level.

This work builds on these observations by exploring a fine-grained and hierarchical model of isolation, where fractions of a machine can be isolated dynamically using migration. Using different units of isolation allows a system to isolate processes from each other with a minimum of over-allocated resources, and having a dynamic and reconfigurable model enables isolation to be procured on-demand. The model is then realised as an implemented framework that allows the fine-grained provisioning of units of computation, managing migrations at the core, virtual CPU, process group, process/container and virtual machine level. Use of this framework is demonstrated in detecting and mitigating a machine-wide covert channel, and in implementing a multi-level moving target defence.

Finally, this work describes the extension of post-copy live migration mechanisms to allow temporary virtual machine migration. This adds the ability to isolate a virtual machine on a short term basis, which subsequently allows migrations to happen at a higher frequency and with fewer redundant memory transfers, and also creates the opportunity of time-sharing a particular physical machine's features amongst a set of tenants' virtual machines.