Quantum Machine Learning Kernel Methods for Robust Anomaly Detection: Understanding Noise Robustness

On May 2, 2025, researchers Kilian Tscharke, Maximilian Wendlinger, Sebastian Issel, and Pascal Debus published a study titled Quantum Support Vector Regression for Robust Anomaly Detection, exploring the application of quantum machine learning in identifying anomalies within large datasets. Their work, conducted under IBM Quantum’s research initiatives, demonstrated that quantum support vector regression (QSVR) achieves strong classification performance across eleven datasets while exhibiting robustness to certain types of noise—though it remains vulnerable to adversarial attacks.

The study explores machine learning approaches for anomaly detection, focusing on kernel methods like QSVR. Benchmarks across eleven datasets show strong performance, outperforming noiseless simulations in two cases. The model demonstrates robustness to certain noises (depolarizing, phase damping) but is vulnerable to others (amplitude damping, miscalibration). Additionally, QSVR proves highly susceptible to adversarial attacks, with noise failing to enhance its resilience.

Recent advancements in quantum computing have unlocked new possibilities for tackling complex challenges in machine learning. Among these innovations, anomaly detection—a critical task across industries such as cybersecurity and healthcare—has seen significant progress using quantum algorithms. Advances in machines such as IBM’s 27-qubit quantum hardware evaluate how these models perform under realistic noise conditions and assess their vulnerability to adversarial attacks. The insights gained from this research highlight the practical challenges and point towards opportunities for deploying quantum machine learning (QML) models in real-world applications during the NISQ era.

The study centres on a quantum support vector regressor for semi-supervised anomaly detection. Six distinct noise channels were applied across multiple datasets to simulate various noise sources. These included amplitude damping, phase damping, depolarising noise, and miscalibration errors—common issues in quantum systems. The model’s resilience against adversarial attacks was also evaluated using Projected Gradient Descent (PGD) attacks of varying strengths.

The results demonstrated that QSVR models maintain a high level of performance even in noisy environments. The average Area Under the Curve (AUC) score on IBM’s quantum hardware was 0.72, compared to 0.76 in noise-free simulations. This suggests that quantum models can deliver results comparable to classical methods despite the challenges posed by noise. However, certain noise types had a more pronounced effect, with amplitude damping and miscalibration errors proving particularly challenging.

The model’s response to adversarial attacks also revealed important insights. While the QSVR demonstrated resilience against weaker attacks, more substantial perturbations significantly degraded performance. This highlights the need for further research into robustifying quantum models against such threats.