Every digital operation on a chip leaves a physical footprint. When logic gates flip, they draw current from the chip's power distribution network. Because power delivery paths have non-zero resistance and capacitance, this sudden current draw induces a localized drop in supply voltage, known as an $IR$ drop.
Historically, observing these voltage fluctuations required physical access to the hardware using external oscilloscopes or probes. The core insight driving this research direction is that we can bypass physical probes entirely by exploiting a fundamental property of semiconductors: transistor switching speed is directly dependent on supply voltage. When voltage drops, transistors take longer to charge and discharge, increasing the propagation delay of signals passing through them. Consequently, internal power fluctuations are naturally transduced into timing variations.
Measuring Voltage Digitally via Delay Lines
To capture these sub-nanosecond voltage fluctuations without external hardware, researchers embed digital structures that act as internal voltmeters. The most common mechanism is the Time-to-Digital Converter (TDC) or a routing delay sensor.
A TDC consists of a series of delay elements (such as buffers) arranged in a chain. A pulse is launched into the chain, and after a fixed time interval (typically one clock cycle), a set of flip-flops samples how far the pulse has traveled. If the local supply voltage is high, the gates switch rapidly, and the signal travels far down the chain. If a concurrent workload draws heavy power, the resulting voltage drop slows the signal down, causing it to cover fewer stages. The digital output of the TDC directly mirrors the real-time voltage profile of the chip.
As explored by Ognjen Glamocanin et al. in Instruction-Level Power Side-Channel Leakage Evaluation of Soft-Core CPUs on Shared FPGAs, these integrated sensors allow for high-resolution power side-channel measurements entirely on-chip, eliminating the need for external equipment.
Exploiting the Delay Signal for Remote Attacks
In shared hardware environments, such as multi-tenant cloud FPGAs or system-on-chips (SoCs), this shared power distribution network becomes a critical security vulnerability. An attacker co-located on the same fabric can deploy a TDC to spy on a neighbor's computational state.
Because different operations and data values generate unique power signatures, the resulting propagation delays can reveal highly sensitive information. For example, in Mercury: An Automated Remote Side-channel Attack to Nvidia Deep Learning Accelerator, Xiaobei Yan et al. demonstrated that an automated sequence-to-sequence deep learning model can translate these on-chip voltage leakage patterns to reconstruct deep learning architectures with minimal error from a single inference run.
Beyond passive eavesdropping, these delay signatures enable highly synchronized active attacks. In DeepStrike: Remotely-Guided Fault Injection Attacks on DNN Accelerator in Cloud-FPGA, Yukui Luo et al. used TDCs to track signal propagation delays to identify specific execution phases or neural network layers through their unique power signatures. This precise timing allows attackers to coordinate fault injections. Building on this concept, Adnan Siraj Rakin et al. developed Deep-Dup: An Adversarial Weight Duplication Attack Framework to Crush Deep Neural Network in Multi-Tenant FPGA, which uses malicious power-plundering circuits to intentionally drop the voltage at the exact moment of a data transfer, inducing surgical "weight duplication" errors that degrade model accuracy.
Harnessing the Side-Channel for Defense and Testing
The same physical phenomenon that enables remote attacks can be repurposed as a zero-latency, non-intrusive defense mechanism. Because delay-based sensing monitors the physical side-channel rather than modifying the software or model architecture, it provides real-time security without impacting system throughput.
In UniGuard: A Unified Hardware-oriented Threat Detector for FPGA-based AI Accelerators, Xiaobei Yan et al. utilized a TDC as a passive voltage drop sensor to detect adversarial, backdoor, and model extraction attacks by identifying anomalous power fluctuations during inference. Similarly, Vincent Meyers et al. demonstrated in Towards Functional Safety of Neural Network Hardware Accelerators: Concurrent Out-of-Distribution Detection in Hardware Using Power Side-Channel Analysis that routing delay sensors can act as digital voltmeters to capture fine-grained voltage drops. By running a score evaluation function independently on the physical side-channel data stream, the authors achieved concurrent out-of-distribution detection with zero latency overhead on the primary task.
This paradigm of using power supply signatures as a proxy for internal circuit behavior also extends to hardware manufacturing and testing. Swarup Bhunia et al. showed in Frequency Specification Testing of Analog Filters Using Wavelet Transform of Dynamic Supply Current that monitoring dynamic supply current fluctuations captures embedded frequency response characteristics, allowing for the non-invasive testing and trimming of analog filters without requiring dedicated output pins.
Macro-Scale Transduction via Frequency Throttling
While TDCs capture micro-level propagation delays within clock cycles, the power-to-timing transduction principle also manifests at a macro scale through system-level power management. Modern processors feature Dynamic Voltage and Frequency Scaling (DVFS) to keep chips within safe thermal and power envelopes.
In Frequency Throttling Side-Channel Attack, Abhishek A. Chakraborty et al. demonstrated that this reactive power management inadvertently creates a software-measurable timing side channel. When a processor processes data that causes higher power consumption, it may hit its power or thermal limits, prompting the hardware to throttle its operating frequency. This converts internal, data-dependent power variations into observable differences in execution time.
Crucially, the authors showed that an attacker does not need administrative control to exploit this. A privileged attacker can lower hardware consumption thresholds to force throttling, while an unprivileged attacker can run high-power workloads on an adjacent core to exhaust the shared system-on-chip power budget, forcing the hardware to throttle the victim's core and exposing its internal state through timing variations.
Key Foundational Work
Deep-Dup: An Adversarial Weight Duplication Attack Framework to Crush Deep Neural Network in Multi-Tenant FPGA (2020) — contributed 1 concept elements to this direction.
Frequency Specification Testing of Analog Filters Using Wavelet Transform of Dynamic Supply Current (2005) — contributed 1 concept elements to this direction.