arXiv:2609.19090
Securing quantum error correction against misleading advice from AI agents
A. Barış Özgüler
Abstract
Can an attacker turn influence over an artificial intelligence (AI) adviser into a harmful quantum error-correction update? We identify an ambiguity in passive syndrome records that obstructs recovery selection, then show how additional calibration measurements support certified recovery updates under uncertainty and drift. In an odd-distance square toric code with error-free preparation, syndrome measurements, and recovery operations, opposite coherent $X$ rotations produce identical passive syndrome-history distributions. Yet a fixed phase correction can help at one sign and harm at the other. A terminal logical measurement on known encoded calibration states supplies the missing sign information. A separate evaluator accepts an update only when calibration uncertainty and a justified drift bound certify improvement over the current recovery, without assuming that the adviser recommends correctly. In simulated advice attacks, calibration-confidence checks reject harmful proposals while retaining beneficial updates under honest advice. We derive sufficient limits on calibration age that require improvement through deployment. In matched simulations, a validated channel-specific bound retains more beneficial updates than the general bound after accounting for evaluation time, while preventing the tested harmful activations under the stated drift assumption. A separate surface-code experiment includes stochastic circuit faults and noise changing during acquisition. Deterministic controllers achieve at least as many beneficial updates with the same observations. Violating the drift assumption permits harmful acceptance in the toric experiment. The results identify information required for recovery selection, establish conditional guarantees against harmful updates, and quantify the recovery improvements forgone through conservative acceptance.
AI Summary
The paper asks when recovery updates proposed by an untrusted or mistaken AI adviser can be certified to improve QEC despite calibration uncertainty and noise drift. It combines an analytical observation obstruction with calibration-based acceptance rules and simulated advice attacks. For an odd-distance square toric code under uniform coherent $X$ rotations, ideal syndrome extraction and deterministic minimum-weight recovery, opposite angles $\pm\theta$ give identical distributions of every passive syndrome history. Syndrome effects depend only on logical product parity; input distinguishability begins at order $\theta^{2L}$ and scales as $\Theta_L(|\theta|^L)$ over the coherent-accumulation horizon $R=\Theta_L(|\theta|^{-L})$. Nevertheless, sign matters for recovery: at $L=3$, $|\theta|=0.10$ and 300 rounds, entanglement infidelity is approximately $0.1324$ without phase correction, $0.000746$ with the correctly signed table, and $0.4140$ with the wrong sign. Terminal logical measurements on known encoded calibration states supply sign information. A separate evaluator accepts only if a confidence region, propagated through acquisition-to-deployment drift, certifies improvement over the incumbent. Valid simultaneous coverage and risk enclosures give a union-bound guarantee for arbitrary adaptive proposals. In 12 toric attack workloads, misleading advice induces 12 harmful proposals, all rejected. A sharper validated drift bound retains 12 delayed beneficial updates versus zero with the general bound. Separate distance-3/5 surface-code simulations include stochastic circuit faults and drift during acquisition. Guarantees depend on trusted measurement and enforcement, valid physical models and drift bounds; violating drift permits harmful acceptance. Timing bounds can reject beneficial updates. Deterministic controllers achieve at least as many beneficial updates with the same observations. No hardware deployment or large-distance scaling is demonstrated.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 74 pages, 32 figures (10-page main text, 62 pages of Supplemental Material, and 2 pages of references)