arXiv:2610.01777
QUFIG: GNN-Based Prediction of Quantum Fault Injection Vulnerabilities with Gate-Level Precision
Shihan Zhao, Qiying Li, Ben Dong, Qian Wang, Yuntao Liu
Abstract
The growing scale and accessibility of quantum hardware exposed new reliability and security challenges in the quantum computing workflow, such as the run-time fault injection attacks in cloud-based quantum computing platforms. However, existing works fail to identify vulnerabilities with gate-level precision or adapt to run-time environments. In this work, we formulate gate-level fault analysis as a learning-guided prioritization problem under restricted fidelity budgets. The framework uses a circuit-DAG-based GNN backbone to predict the vulnerability score of each gate to each type of injected fault, defined as the impact of the gate-fault pair on circuit fidelity. The gate-fault pairs are then ranked by their vulnerability score. Experiments on QASMbench and HamLib MaxCut show that QUFIG recovers high-impact vulnerable gate-fault pairs with fewer inspections than random and depth-based heuristics. Our results show that QUFIG can reduce the number of gates requiring inspection by 2.9--19.8% while maintaining effective fault identification, allowing quantum circuit designers to identify vulnerabilities and apply targeted defenses more efficiently.
AI Summary
QUFIG addresses how to prioritize gate-level fault injections when evaluating every gate–fault pair is too costly. It learns to rank candidate injections by their predicted impact on circuit outputs, targeting budget-constrained simulation and security auditing rather than exhaustive fault analysis. The framework represents compiled circuits as directed acyclic graphs with gate-type, position, qubit-index, and parameterization features. Standard GNN backbones produce fault-specific probabilities that a gate–fault pair belongs to a circuit’s highest-impact subset. Training uses simulated fault-injection labels and a combined binary-cross-entropy and mean-absolute-error loss. Four channels are considered: amplitude damping, phase damping, bit flip, and depolarizing noise. Impact is defined using output-distribution fidelity loss, optionally combined with normalized application-objective deviation. Evaluation uses QASMbench, with 35 original circuits expanded to 2,907 through concatenation, and 2,120 HamLib MaxCut circuits. Group-aware splits exclude related circuit structures from training when testing. GraphSAGE achieves AUPRC values of 0.220 and 0.283, respectively. For the reported top-16% vulnerable-gate recovery target, it requires inspecting 56.3% of MaxCut gates, versus 70.1% for random selection and 69.3% for reverse-depth traversal. Corresponding QASMbench costs are 70.6%, 75.8%, and 72.7%. At the top-20% QASMbench target, inspection falls from 90.9% to 81.9% relative to random selection. The study is simulation-based and restricted to the specified noise channels and circuit suites. It notes limited circuit diversity and smaller improvements on QASMbench, attributed to stronger depth-related structure. The framework uses existing message-passing operators rather than introducing a new GNN architecture.
Central claim may not hold: QUFIG predicts probabilities of belonging to a high-impact subset rather than fidelity-loss magnitudes, without establishing equivalence to its stated objective of maximizing total selected fault impact. The main results count inspected gates although the method ranks gate–fault pairs, and the paper does not define score aggregation or the fault-channel evaluation cost per inspected gate. The headline inspection savings depend on a recovery criterion and aggregation protocol that are insufficiently specified to reconstruct what level of vulnerable-gate recovery the reported budgets achieve.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Accepted by IEEE International Conference on Computer Design (ICCD) 2026