PaperlayerView exact revision on arXiv

Source-linked AI summary

Hardware Fingerprinting FTQC via Quantum Decoder Timing

Friedrich Doku, Jakub Szefer, Kaitlin N. Smith

arXiv:2609.12145v1cs.CR

TL;DR

FTQC decoders must process syndrome rounds in real time, yet their data-dependent wall-clock latency can expose information to a passive observer. This work uses per-shot timing to recover workload and device information, estimating p_L within a factor of two on 84–94% of runs and fingerprinting devices with up to 89% accuracy.

  • Problem

    Decoder latency sits on FTQC’s real-time critical path, and whether its wall-clock timing leaks workload information is an unresolved security question.

    Source: §I · p. 1

  • Method

    The study analyzes per-shot wall-clock latencies from data-dependent decoders to infer detector counts, code distance, and device identity.

    Source: §VIII · p. 11

  • Results

    Timing estimates p_L within a factor of two on 84–94% of runs, infers code distance with 95% accuracy from as few as 50 timed shots, and fingerprints processors with up to 89% accuracy.

    Source: §VIII · p. 11

  • Takeaways & Limitations

    Decoder wall-clock execution time constitutes an exploitable side channel from which a passive observer can recover sensitive workload information.

    Source: §VIII · p. 11

    Evaluation covers three IBM Heron devices and two code distances, without testing how classification accuracy scales as the candidate-device pool grows.

    Source: §VI · p. 10

Abstract

from arXiv · show

As the quantum computing field transitions toward Fault-Tolerant Quantum Computing (FTQC), intensive efforts are focused on scaling architectures and realizing active error correction. However, this shift introduces security surfaces that remain largely unexplored. Fault-tolerant quantum computers pair a quantum processor with a classical decoder that sits on the critical path of every syndrome-extraction round. For the first time, this work demonstrates that the wall-clock time each decoder takes to process a syndrome measurement and decoding round constitutes a novel, exploitable hardware side channel on physical quantum hardware. Using per-shot decoder timings from three IBM Heron processors collected over a 68-day window, the decode-time distribution alone allows a passive observer to (i) reconstruct the shot-by-shot detector-firing distribution and estimate the workload's logical error rate $p_L$, (ii) infer the code distance in use, and (iii) fingerprint the specific physical device with up to 89% accuracy (a random guess is 33%), with a pooled two-sample Kolmogorov-Smirnov test confirming the decode-time distributions are statistically distinct. In noisy simulation inspired by public data from Google's 105-qubit Willow processor, decoder timing further distinguishes 9 surface-code patches at different locations on the chip with 81% accuracy, showing the side channel persists on below-threshold fault-tolerant hardware from a different vendor and code family.

I. INTRODUCTION

The introduction frames decoder latency as both a real-time constraint and an information leak: per-shot timing can expose workload properties and identify the physical device, even without syndrome access.

I. INTRODUCTION

  • FTQC requires a classical decoder to return corrections before the next syndrome round; falling behind lets uncompensated syndrome data accumulate and can stall logical computation.Decoder latency is therefore a load-bearing constraint of the quantum stack.Source: §I · p. 1

A. The Observation

  • A denser syndrome takes longer to decode, and device-specific error characteristics shape a per-shot timing distribution that can identify the hardware.The passage links detector firings to physical error rates, including gate infidelities, T1/T2 heterogeneity, correlated errors, and readout asymmetry.Source: A. The Observation · p. 1

B. The Stakes

  • In shared quantum-cloud queues, identifying the processor can confirm co-residency with a target and defeat provider anonymization.Prior hardware-fingerprinting concerns on classical shared infrastructure also connect device identification to cross-tenant data exfiltration.Source: B. The Stakes · p. 2

C. Attacks on Real-hardware

  • The evaluation used identical repetition-code memory circuits on three IBM Heron processors over 68 days, recording per-shot MWPM and Union-Find decoder latencies.The timing analysis was complemented by surface-code data from Google’s 105-qubit Willow processor.Source: C. Attacks on Real-hardware · p. 2
  • Timing reconstructs detector-firing counts, separates d = 5 from d = 7, and identifies the physical device with up to 89% accuracy versus one-third for random guessing.A pooled two-sample Kolmogorov–Smirnov test found every device pair had distinct timing distributions; Willow surface-code patches were also distinguished by decoder timing.Source: C. Attacks on Real-hardware · p. 2

II. THREAT MODEL

The threat model gives a passive observer access only to per-shot decoder execution times, supplemented by a modest labeled reference baseline. The observer neither interacts with the quantum computation nor sees its syndrome bits or workload identity.

II. THREAT MODEL

  • The attack surface is a data-dependent decoder’s per-shot wall-clock latency, which the observer can obtain through channels such as cloud contention, exposed telemetry, or physical side channels.The assumed decoder processes each syndrome and returns a correction; denser syndromes take MWPM and Union-Find longer to resolve.Source: §II · p. 2

A. Attacker Capabilities

  • The passive, honest-but-curious adversary observes decoder timing only and does not tamper, inject faults, interact with hardware, or see syndrome bits or workload identity.Every result is reconstructed solely from decode timing.Source: A. Attacker Capabilities · pp. 2–3

B. Attacker Knowledge

  • The observer has a modest labeled baseline of prior decode-time distributions for firing count, code distance, or device, which implicitly encodes the decoder family and code.The baseline is described as cheap to acquire and durable.Source: B. Attacker Knowledge · p. 3

C. Attacker Objective

A passive observer uses per-shot decoder times and a reference baseline to reconstruct detector-firing counts, estimate logical error rate, infer code distance, and fingerprint the physical device. The attack combines firing-count probabilities with decode-time statistics because timing reflects both detector activity and device noise.

C. Attacker Objective

  • The attacker uses per-shot decode times and a reference baseline to infer firing counts and estimate pL, infer code distance d, and identify the physical device.Together, these inferences profile which device is running which workload and how well, without access to syndrome bits.Source: C. Attacker Objective · p. 3

III. METHODOLOGY

  • The pipeline extracts firing-count probabilities and decode-time summary statistics, combines them into a 14-dimensional fingerprint, and compares it with a labeled reference bank using k-nearest neighbors.The reference bank contains previously observed fingerprints; k-nearest neighbors assigns the query run to a device.Source: §III · p. 3, C. Hardware Fingerprinting · p. 3

A. Decoder Latency as a Feature Source

  • Per-shot decoder latency rises with fired-detector count, making it a real-time proxy for the device’s noise profile; differing noise profiles can produce differing latencies.The timing-to-firing-count mapping is a property of the decoder, while device noise profiles affect the observed latencies.Source: A. Decoder Latency as a Feature Source · p. 3

B. Code Distance and Detectors

  • Larger code distance d shifts decode-time distributions toward longer times and heavier tails, allowing an observer with reference distributions to distinguish candidate distances.Because operators choose d to target a logical failure rate, its disclosure also reveals the workload’s intended fidelity regime.Source: B. Code Distance and Detectors · p. 3

C. Hardware Fingerprinting

  • The 14-dimensional fingerprint combines nine firing-count probabilities, including a tail bucket for counts of at least eight, with five decode-time statistics.The timing statistics are the mean, standard deviation, log-mean, 95th percentile, and 99th percentile; together the feature families represent error volume and runtime distribution.Source: C. Hardware Fingerprinting · p. 3
  • Cross-validation found the 30-bin time histogram redundant, while removing firing probabilities reduced d=7 accuracy and removing timing statistics impaired d=5 accuracy.The authors report that the resulting 14 features are the minimal set needed to succeed across both code distances.Source: C. Hardware Fingerprinting · p. 4

D. Classification

The classifier compares a query fingerprint with labeled fingerprints in a reference bank to predict the device. The data come from three Heron-R2 processors whose daily calibration drift is treated as part of the fingerprinting problem.

D. Classification

  • A k-nearest-neighbor classifier predicts a query run’s device by comparing its fingerprint with labeled entries in a growing reference bank.The authors chose k-NN as an interpretable, stable baseline that can add fingerprints without retraining.Source: D. Classification · p. 4

A. Devices and Calibration

  • The corpus uses three IBM Heron-R2 processors with the same architecture but different physical realizations and daily recalibrations.Calibration drift is not controlled; the authors treat it as part of the problem and note that excessive drift can make one device resemble another.Source: A. Devices and Calibration · p. 4

B. Circuit: Repetition-Code Memory Experiment

The study runs a repeated-syndrome-extraction repetition-code memory circuit on IBM Heron and times decoding locally under a controlled decoder setup. It uses timed shots to reconstruct firing-count information, while retaining ground-truth counts only for validation.

B. Circuit: Repetition-Code Memory Experiment

  • Each repetition-code shot uses d data qubits and d−1 ancillas, performs five syndrome-extraction rounds around a transversal logical X, and is expected to yield logical outcome |1⟩.A shot is counted as a logical error when the decoder’s correction disagrees with that expected outcome; using the same circuit across devices attributes timing differences to hardware rather than workload.Source: B. Circuit: Repetition-Code Memory… · p. 4

C. Timing Instrumentation

  • Decodes run locally on one isolated CPU core, with nanosecond-resolution timing and decoder-specific instrumentation to limit timing overhead.MWPM is timed inside a custom C extension; Union-Find is timed in Python after cache-warming decodes.Source: C. Timing Instrumentation · pp. 4–5

D. Dataset

  • Across 14 runs over 68 days, each run-device-distance combination supplies 48,000 shots for pL and 16,000 timed shots with ground-truth detector counts reserved for validation.The attack reconstructs firing counts by mapping each shot’s log decode time to the nearest reference mean for a firing count; true counts are not attack features.Source: D. Dataset · p. 5
  • The study times MWPM (via PyMatching) and Union-Find on an isolated CPU core; only MWPM is specified to use a uniform error model, p = 10^-3, without per-device calibrations.That same uncalibrated MWPM instance was applied across all three devices, so timing-distribution differences were attributed to raw hardware error patterns rather than decoder configuration.Source: D. Dataset · p. 5

V. EVALUATION

The evaluation asks whether timing alone can reveal detector-firing distributions and pL, code distance, and device identity. For detector-count inference, both decoders show latency increasing with fired-detector count, although the separation is stronger for Union-Find.

V. EVALUATION

  • The evaluation tests whether decoder timing can reconstruct firing counts and estimate pL, infer code distance, and fingerprint hardware across days.The authors also describe a subsequent test of fingerprinting on Google Willow data to assess transfer to below-threshold surface-code hardware.Source: §V · p. 5

A. Detector-Count Inference from Decode Times

  • Union-Find’s per-shot times form distinct, progressively longer bands by firing count, while MWPM’s bands overlap more; both show mean latency increasing with fired-detector count.The plotted IBM Fez data cover d=5 and d=7; the figure omits zero-firing shots, which decode almost instantly.Source: Figure 4 · p. 5, A. Detector-Count Inference from… · p. 6
  • Both decoders’ timing reflects input-dependent work, but Union-Find’s runtime is more sensitive to firing count while MWPM’s precomputation flattens its slope.The authors attribute the difference to Union-Find’s per-detector work and MWPM’s largely precomputed work with smaller per-firing updates.Source: A. Detector-Count Inference from… · p. 6
  • For both MWPM and Union-Find at d=7, mean per-shot decode time grows linearly with fired-detector count, with each decoder shown on its own scale.The figure cautions against comparing absolute decoder latencies because Union-Find is a reference Python implementation and MWPM is PyMatching.Source: Figure 5 · p. 6

B. Logical Error Rate

Decoder timing provides a factor-of-two estimate of logical error rate more often than a timing-free median baseline, using per-distance log-linear models of estimated detector-firing counts. The estimates remain predictive across runs on three IBM Heron processors over 68 days.

B. Logical Error Rate

  • The evaluation used 14 runs across three IBM Heron devices over 68 days, comparing timing-derived firing counts with pL measured from 48,000 shots per run-distance-device cell.Ground-truth pL was the fraction of shots where the decoder's correction failed to restore the circuit's expected logical outcome.Source: B. Logical Error Rate · p. 6
  • The per-distance log-linear models estimate pL from the mean firing count inferred from decoder timings.Taking log10 pL as the response makes the observed nonlinear relationship with firing count linear for ordinary least-squares fitting; each additional average fired detector multiplies pL by 10^a.Source: B. Logical Error Rate · p. 6
  • 94% of d=5 and 84% of d=7 runs were predicted within a factor of two of the true pL, versus baseline rates of 59% and 25%.Figure 6 plots timing-estimated mean firing count against measured pL for individual runs, with separate log-linear fits by distance.Source: Figure 6 · p. 6, B. Logical Error Rate · p. 7

C. Code-Distance Discrimination

Batched decoder timings distinguish code distance, but the required shot count depends on the decoder and distance inference needs a fresh, device-specific noise reference. In contrast, converting timing to firing counts uses a baseline that transfers well across devices and time.

C. Code-Distance Discrimination

  • For 16,000 IBM Fez shots per distance, decode-time distributions differed significantly between d=5 and d=7 for MWPM (D = 0.178, p < 10^-221) and Union-Find (D = 0.401, p < 10^-300).The higher distance shifts timing distributions toward longer latencies, but per-shot mean shifts are smaller for MWPM (Cohen’s d = 0.10) than Union-Find (0.56).Source: C. Code-Distance Discrimination · p. 7
  • 95% accuracy requires about 35 Union-Find shots or 122 MWPM shots under a same-run reference; accuracy exceeds 99% by about 250 shots.Figure 8 plots accuracy against the number of observed shots; shot-count thresholds are approximate.Source: Figure 8 · p. 7, C. Code-Distance Discrimination · p. 7, D. Baselines · p. 7

D. Baselines

  • The timing-to-firing-count baseline transfers across Heron devices and time: mean latency varies by a median under 4% across devices, while drift on Fez can reach about 25% on occasional recalibration days.A batch recovers the firing-count distribution despite overlap between adjacent firing-count timing distributions.Source: D. Baselines · pp. 7–8
  • Distance inference needs a batch and a reasonably fresh, device-specific noise baseline because physical noise drifts and shifts the firing-count distributions.The threat model assumes periodic refresh using at most one known-distance run per device near calibration; the distance signal is a systematic offset at the same noise level.Source: C. Code-Distance Discrimination · p. 7, D. Baselines · p. 8

E. Hardware Fingerprinting

Decode-time fingerprints cluster by physical processor in a visualization of standardized run features, and pairwise tests confirm that device timing distributions differ. Separation is statistically significant across every device pair, code distance, and decoder tested.

E. Hardware Fingerprinting

  • Runs from the same processor form visually distinct clusters in every code-distance and decoder panel of the 2D MDS projection.The projection visualizes distances between standardized 14-dimensional fingerprints; the classifier uses the raw feature vectors rather than the lossy projection.Source: E. Hardware Fingerprinting · p. 8, Figure 9 · p. 8

1) KS Testing:

  • The fingerprinting test holds code distance constant and compares decode-time distributions across physical machines using pairwise KS tests.D is the maximum distance between the distributions' cumulative distribution functions.Source: 1) KS Testing · p. 8
  • Every pair of IBM Heron devices has statistically distinct decode-time distributions at p < 10^-100, with the largest reported separation D = 0.425 for Fez versus Kingston at d=7 under MWPM.Table II reports pairwise Kolmogorov–Smirnov statistics; larger D indicates more distinct distributions.Source: Table II · p. 8, 1) KS Testing · p. 8

2) Device Identification:

Decoder timing identifies IBM Heron devices with 81–89% accuracy and distinguishes surface-code patches on Google Willow. On Willow, timing-only patch identification reaches 81% for nine distance-3 patches, with further patch-identification results at distance 5.

2) Device Identification:

  • The 14-dimensional fingerprint combines firing-count probabilities and decode-time tail statistics, classified by the majority label among the three nearest fingerprints.Identifiability tracks stable spatial noise patterns: Kingston has one consistently hot chain position, Fez is hot at both ends, and Marrakesh lacks a stable hotspot.Source: 2) Device Identification · pp. 8–9
  • Timing alone identifies the correct IBM Heron device in 81–89% of runs, versus 33.3% random guessing; Union-Find reaches 89.2%.Fez and Kingston are usually identified above 90%, while Marrakesh is hardest at 71–79%.Source: Table III · p. 9, 2) Device Identification · p. 9

F. Generalization to Surface Codes: Google Willow

  • On Google Willow, timing identifies the source of nine distance-3 patches with 81% accuracy, against 11% random guessing.Each patch contributes 20 batch fingerprints; the points cluster by patch in the L1 MDS figure. The dataset is a single-day snapshot, unlike the multi-calibration Heron evaluation.Source: F. Generalization to Surface Codes:… · p. 9, Table IV · p. 9, Figure 10 · p. 9
  • Willow decoding uses physical detection events with one fixed MWPM decoder and a uniform SI1000 noise model; mean decode time rises from 10.4 µs at d=3 to 40.7 µs at d=5 and 92.3 µs at d=7.Timing identifies the nine d=3 patches at 81% and four d=5 patches at 82%; the single d=7 patch leaves no patch-identification task.Source: F. Generalization to Surface Codes:… · pp. 9–10, Table IV · p. 9

VI. DISCUSSION

The discussion assumes a passive observer can measure per-shot latency from a data-dependent decoder, then considers defenses that remove this dependence or hide timing.

VI. DISCUSSION

  • The threat model assumes a passive observer can measure per-shot latency from a data-dependent decoder, so defenses must remove data dependence or prevent latency observation.The discussion considers both changing decoder behavior and restricting access to timing.Source: §VI · p. 10
  • Constant-time decoding closes the channel by construction, but latency padding throttles throughput to worst-case speeds and fixed-latency hardware adds complexity.Both approaches trade security against hardware complexity and real-time processing rates.Source: §VI · p. 10
  • Shot batching hides granular per-shot latency but can still expose mean detector-firing rates; random delays can be averaged away with more shots.The text identifies strict architectural isolation from execution-time observation as necessary to fundamentally prevent these timing attacks.Source: §VI · p. 10
  • The evaluation uses identical circuits across devices, three IBM Heron processors, two code distances, and MWPM and Union-Find, leaving workload effects, larger device pools, and qLDPC decoders untested.The authors leave disentangling workload-specific effects from device-specific noise and characterizing iterative qLDPC decoders such as BP+OSD to future work.Source: §VI · p. 10

VII. RELATED WORK

The paper distinguishes decoder timing from prior quantum side channels and hardware-fingerprinting approaches, presenting it as a passive signal that can reveal device and workload information.

A. Side-Channel Attacks on Quantum Computers

  • Prior quantum side channels target control-stack traces, total circuit execution time, or active crosstalk; this work instead observes passive per-shot classical decoder latency.The paper describes its decoder channel as specific to fault-tolerant workloads.Source: A. Side-Channel Attacks on Quantum… · p. 10

B. Quantum Hardware Fingerprinting

  • Earlier fingerprinting uses directly readable hardware features or syndrome strings, including a 99%-accurate syndrome method [22], while decoder timing identifies devices up to 89% without those observables.The timing-based attacker does not observe syndrome bits, calibration data, or the circuit.Source: B. Quantum Hardware Fingerprinting · p. 10

C. Decoder Latency and Real-Time Decoding

  • Decoder-runtime variance has been studied as a real-time performance bottleneck; this work demonstrates that the same variance is an exploitable security channel and characterizes its leaks.Prior responses include hardware designs that bound worst-case latency and elastic decoders to manage backlogs.Source: C. Decoder Latency and Real-Time… · p. 10

VIII. CONCLUSION

  • Timing alone estimates pL within a factor of two on 84–94% of runs, infers code distance with 95% accuracy from as few as 50 timed shots, and fingerprints processors with up to 89% accuracy.Willow-inspired experiments confirm fingerprinting generalizes to surface codes.Source: §VIII · p. 11
Loading 2609.12145v1…