Adversarial Robustness of Partitioned Quantum Classifiers

Pouya Kananian, Hans-Arno JacobsenarXiv 2025 · DOI · arXiv

Abstract

As quantum computing transitions into the Noisy Intermediate-Scale Quantum (NISQ) era, partitioning large quantum circuits across multiple devices has become a necessary strategy to overcome hardware qubit limitations. However, this distributed approach introduces novel security vulnerabilities. This paper investigates the adversarial robustness of partitioned quantum classifiers, demonstrating that the processes of circuit cutting and quantum teleportation create new attack surfaces. We prove that tampering with these internal communication channels is equivalent to inserting malicious unitary gates within the circuit's intermediate layers. Our findings reveal that these internal attacks can be more disruptive than traditional input-level perturbations, significantly degrading classification accuracy even with localized, low-strength attacks.

Introduction: The Qubit Gap and Distributed Quantum Computing

A significant hurdle in current quantum machine learning (QML) is the "qubit gap"—the discrepancy between the high qubit counts required by complex algorithms and the limited capacity of individual NISQ chips. To bridge this gap, researchers employ techniques like circuit cutting (decomposing a circuit into smaller fragments connected by classical communication) and quantum teleportation (linking fragments via quantum channels). While these methods enable the execution of larger models, they often operate under the assumption of secure communication and state preparation. This research challenges that assumption by exploring the impact of adversarial tampering within these partitioned architectures.

Modeling Adversarial Attacks in Partitioned Circuits

The core contribution of this work is a formal framework for analyzing "partitioned quantum classifiers." Unlike traditional adversarial QML, which primarily focuses on perturbing input states, this study examines internal robustness.

The process of circuit cutting and its functional equivalence to inserting an adversarial gate into the circuit's internal layers.

The authors demonstrate that manipulating the "wire cutting" or "teleportation" steps in a distributed system is mathematically identical to inserting adversarial unitary perturbations at various depths of the circuit. This mapping allows for a rigorous study of how internal vulnerabilities affect the final classification outcome.

Theoretical Framework and Bounds

To quantify the impact of these attacks, the paper establishes two key theoretical results:

  • Theorem 6.1: Provides an upper bound on the "predictive confidence shift"—the change in output probability—based on the Diamond Distance and Operator Norm of the adversarial perturbations.
  • Theorem 6.3: Utilizes Chebyshev’s inequality to derive a probabilistic bound for Haar-random inputs. This allows researchers to estimate the likelihood of an attack flipping a classification label based on the "attack strength" (the distance of the adversarial gate from the identity operator).

These bounds are shown to be particularly tight for "stealthy" attacks involving low-strength perturbations, making them reliable tools for predicting model disruption.

Key Insights: The Vulnerability of Depth

The research highlights several critical insights regarding the security of distributed quantum systems:

Architectural configurations of quantum classifiers: (a) a standard unperturbed classifier, (b) a partitioned classifier with a single internal adversarial gate, and (c) a multi-layer partitioned classifier with multiple adversarial gates.
  • Expanded Attack Surface: Partitioned classifiers are inherently more vulnerable than their monolithic counterparts. While a standard classifier is typically only susceptible at the input stage, a partitioned one can be compromised at every "cut" or "teleportation" point.
  • Potency of Internal Gates: Adversarial gates planted in intermediate layers can produce more significant performance degradation than input-level manipulations.
  • Strategic Local Attacks: Even "local" adversarial gates affecting only a few qubits can be highly effective if placed strategically within the circuit architecture.

Experimental Results

The theoretical framework was validated through simulations using PennyLane and Keras on standard benchmarks: MNIST, Fashion-MNIST, and CIFAR-10. The results were stark: classifiers that achieved nearly 100% accuracy on MNIST saw their performance plummet as attack strength increased. The experiments also confirmed that while deeper models might offer slightly better resilience in some cases, they remain highly susceptible to coordinated multi-layer attacks.

Conclusion and Future Outlook

This study reveals a critical security trade-off in the quest to scale quantum computing through partitioning. While distributed architectures solve the qubit gap, they introduce significant internal vulnerabilities. Future research must address the transition from noise-free simulations to real NISQ hardware, where inherent quantum noise might provide a natural defense. Furthermore, the development of specialized "adversarial training" techniques will be essential to secure the next generation of distributed quantum machine learning models.

Summary updated

BibTeX
@misc{kananian2025adversarial,
  author = {Pouya Kananian and Hans-Arno Jacobsen},
  title = {{Adversarial Robustness of Partitioned Quantum Classifiers}},
  year = {2025},
  eprint = {2502.20403},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.20403},
}
Figure