Noise-Resistant Feature-Aware Attack Detection Using Quantum Machine Learning

TL;DR

A quantum machine learning (QML)-based attack detection framework (QML-ADF) that safeguards the security of high-rate CV-QKD systems and will not only enable robust detection of quantum attacks under realistic deployment conditions but also strengthen the practical security of quantum communication systems.

Abstract

Continuous-variable quantum key distribution (CV-QKD) is a quantum communication technology that offers an unconditional security guarantee. However, the practical deployment of CV-QKD systems remains vulnerable to various quantum attacks. In this paper, we propose a quantum machine learning (QML)-based attack detection framework (QML-ADF) that safeguards the security of high-rate CV-QKD systems. In particular, two alternative QML models -- quantum support vector machines (QSVM) and quantum neural networks (QNN) -- are developed to perform noise-resistant and feature-aware attack detection before conventional data postprocessing. Leveraging feature-rich quantum data from Gaussian modulation and homodyne detection, the QML-ADF effectively detects quantum attacks, including both known and unknown types defined by these distinctive features. The results indicate that all twelve distinct QML variants for both QSVM and QNN exhibit remarkable performance in detecting both known and previously undiscovered quantum attacks, with the best-performing QSVM variant outperforming the top QNN counterpart. Furthermore, we systematically evaluate the performance of the QML-ADF under various physically interpretable noise backends, demonstrating its strong robustness and superior detection performance. We anticipate that the QML-ADF will not only enable robust detection of quantum attacks under realistic deployment conditions but also strengthen the practical security of quantum communication systems.

Figures (12)

• Figure 1: CV-QKD scheme using Gaussian-modulated coherent states. Alice’s mode $A$ is coupled to Eve’s mode $e^{\prime}$ via a beam splitter with transmittance $T$, resulting in output modes $B$, received by Bob, and $E$, retained by Eve. GM: Gaussian modulation; Hom: homodyne detection.
• Figure 2: Feature extraction and attack data collection. During Alice's state preparation, a laser diode generates the initial pulses, which are split into a weak signal and a strong LO via a beam splitter (BS). In-phase and quadrature (IQ) modulation is applied on the signal path to prepare coherent states whose quadratures are modulated according to a Gaussian distribution. A delay line is employed on the LO path to synchronize the signal. Polarization beam splitters (PBS) are utilized for multiplexing and demultiplexing. During Bob’s measurement, a delay line on the signal path adjusts the signal timing, whereas the phase modulator (PM) on the LO path allows for random selection of the quadrature to be measured. A power meter (P-Meter) monitors the LO intensity, and a clock ensures precise synchronization. The first homodyne detector measures the received signal, whereas the second performs real-time shot-noise analysis. Polarization controllers optimize the polarization of both the signal and LO to maximize interference efficiency in the homodyne detection. Finally, the collected data ($X_B$, $P_B$, $N_0$, $I_{lo}$) is forwarded to the data processing unit (DPU) for attack detection and raw key distillation.
• Figure 3: QML models for noise-resistant and feature-aware attack detection. (a) QSVM. The core of the model is quantum kernel estimation, which evaluates the fidelity between two quantum feature states. This is implemented by sequentially applying the unitary operation $\mathcal{G}(\vec{x}_i)$, its inverse $\mathcal{G}^\dagger(\vec{x}_j)$, and subsequently measuring all qubits at the output of the quantum circuit. (b) QNN. The model's performance is evaluated using a loss function defined over the trainable parameters $\bar{\theta}$. Optimization is performed via the classical Adam algorithm, which iteratively explores the parameter space defined by an ansatz $\mathcal{J}(\bar{\theta})$. At each iteration, Adam computes the loss with the current parameters and updates them according to adaptive estimates of first and second moments. This process is repeated until convergence to an optimal solution.
• Figure 4: Schematic of variational quantum circuits.
• Figure 5: Performance comparison across different QSVM variants. (a) Mean accuracy comparison of six QSVM variants on the known quantum attack dataset. (b) Mean accuracy comparison of the same six variants on the unknown quantum attack dataset. Each dataset is partitioned into five folds for five-fold cross-validation, where in each iteration, four folds are used for training and one for testing. Accuracy is recorded for each iteration, and the final mean accuracy is computed by averaging the results across all five folds.