2026 Theses Doctoral

Secure Power Management In Emerging Computing Systems

Piersma, Ryan R.

Computer systems, caught in a positive feedback loop between more complex hardware and more demanding applications, have continually required a rethinking of how electrical power is physically distributed among their components. Ensuring that each compute component within a system receives the right amount of power at the right time while meeting system-level objectives such as maximum energy efficiency or performance is a difficult and intricate task. Unfortunately, this ensures that a system's security in relation to its power management is often overlooked despite the fact that security is a full-system property that is not just confined to functional compute resources. This dissertation systematically analyzes how security can be compromised through power management methods in emerging computing systems.

In emerging computing systems, much of the focus with respect to power management has fallen on support for heterogeneous many-accelerator systems-on-chip (SoC) with extremely large numbers of compute cores (up to hundreds of thousands) which are assembled from groups of "chiplets". Power delivery for these systems requires scalability more than even efficiency. Where traditional systems could be controlled by a single entity that made power management decisions for each core, emerging systems distribute the process of power management among the cores, allowing them to perform localized computations that lead to a globalized desirable result. This creates a new surface for interactions between compute nodes that is explored by a number of works in this thesis.

First, I demonstrate an information-theoretic method for understanding information leakage between compute elements in these emerging systems. This method uncovers high bandwidth covert channels and side channels on the order of tens to hundreds of Mbps when a network-on-chip (NoC) is used for compute nodes to pass messages. Following this, I design a new power management scheme called SecureCoin that distributes inter-component messages with randomness, thwarting up to 95% of attempted bit transmissions in order to evade attacker reliance on system topology. SecureCoin also shows significant speedups of up to 2.6x in operation compared to the state-of-the-art.

Next, I present SPIF, a scalable power isolation framework that enables performant power isolation between workloads that share the same physical power source. When emerging compute systems are deployed in multitenant scenarios, the way that power messages are passed between nodes breaks isolation guarantees between workloads. This leads to potential issues with confidentiality, integrity, and availability. SPIF introduces microarchitectural support for power routers to avoid power messages passing through different domains from affecting the state of the routers in other domains and enables more performant distributed power management compared to the state-of-the-art.

This leads to a final observation about decentralized power management, where I show how runtime variability caused by decentralized power management endangers safety-critical workloads by providing an additional source of variability. To address this issue, I present a new power management scheme called StableCoin that enables such systems to meet 70% more difficult deadlines while meeting 94% of deadlines that were broken by previous state-of-the-art systems. This addresses availability and integrity of decentralized hardware power management schemes.

In emerging systems less centered around scalability, chip designers are still making gains in power and energy efficiency by deploying machine learning techniques. As a step toward exploring the vast surface of possible ML-based attacks, I demonstrate model-stealing of power management policies that are implemented as machine learning models with 92% F1 classification score.