Abstract
Latest ARM processors are approaching the computational power of x86 architectures while consuming much less energy. Consequently, cloud providers like Amazon, Microsoft, and Oracle are increasingly offering ARM-based instances. While ARM processors are more energy-efficient, a malicious cloud provider could intentionally reduce the CPU voltage to further lower its costs. This practical experience report describes a novel attack scenario: the Scrooge Attack, launched by a cloud provider against its users to aggressively reduce processor voltage for energy savings. We mimic ARM-based cloud instances using Raspberry Pi devices and demonstrate the degree to which this attack is relevant. The attack is unnoticeable by our detection method up to an offset of -50mV. Finally, we propose a set of client-based detection methods that can identify undervolted instances by pushing them to their thermal limits.
Introduction
ARM processors, such as the AWS Graviton series, have gained significant traction in cloud data centers due to their high performance-per-watt. However, the total control cloud providers exert over hardware and firmware introduces a new vulnerability: intentional undervolting. By operating CPUs below nominal voltage specifications, providers can significantly reduce dynamic power consumption and cooling costs. This "Scrooge Attack" directly benefits the provider's bottom line but compromises the reliability of the user's workloads by eroding the processor's safety guardbands.
The Scrooge Attack Mechanism
In a Scrooge Attack, the provider manipulates firmware configurations to undervolt the CPU. To maintain stealth, the provider must prevent the user from detecting the lowered voltage through standard operating system tools.
As illustrated in the attack workflow, the provider can intercept OS-level requests for voltage readings. By returning nominal values instead of actual undervolted readings, the provider ensures the attack remains invisible to standard monitoring tools. This necessitates a detection strategy that does not rely on software-reported sensors.
Detection via Thermal Guardband Analysis
We propose a detection method that leverages the physical relationship between temperature and circuit stability. As temperature increases, the voltage required for stable operation also increases. By intentionally overheating the CPU, a user can force an undervolted processor to exit its safe operating region.
Our detection tool uses a highly intensive CPU-bound multiplication benchmark to maximize heat dissipation. If the hardware is undervolted, the combination of high temperature and low voltage triggers timing violations, leading to kernel panics. While disruptive, a hard crash provides definitive proof of hardware instability, allowing users to hold providers accountable for Service Level Agreement (SLA) violations.
Experimental Evaluation
We evaluated the Scrooge Attack using Raspberry Pi 3B, 3B+, and 4B models, employing stress-ng for workload simulation and physical power monitors for ground-truth data.
- Profitability: Undervolting yielded energy savings between 5% and 13% on average, reaching up to 37% for specific workloads like
hrtimersunder passive cooling. - Stealth Threshold: The attack remains undetectable by thermal stress up to a voltage offset of -50 mV.
- Detection Success: At more aggressive offsets (e.g., -100 mV), the detection script reliably provoked system failures at 40°C. At milder levels, failures were induced by pushing temperatures to 60°C, with success rates between 40% and 90% depending on the hardware model.
- Latency: The time required to induce a crash via thermal stress ranged from 145 to 250 seconds of sustained load.
Conclusion
The Scrooge Attack demonstrates that cloud providers have a financial incentive to covertly undervolt ARM hardware. While providers can manipulate software-reported telemetry, they cannot hide the physical instability caused by reduced voltage guardbands. Our thermal-based detection method offers a way for users to verify hardware integrity. Future research will extend these findings to enterprise-grade ARM servers and explore more deterministic fault-injection techniques.
Summary updated