[Paper] QUPID: A Partitioned Quantum Neural Network for Anomaly Detection in Smart Grid

Source: arXiv - 2601.11500v1

Overview

The paper introduces QUPID, a partitioned quantum neural network designed to spot anomalies in smart‑grid operations. By marrying quantum‑enhanced feature encoding with a clever workload‑partitioning scheme, the authors claim to beat classical machine‑learning baselines while also being more resistant to adversarial attacks and privacy‑preserving.

Key Contributions

• QUPID architecture: a novel partitioned quantum neural network (PQNN) that splits a large quantum circuit into smaller sub‑circuits, enabling scalable training on near‑term quantum hardware.
• R‑QUPID extension: integrates differential privacy (DP) into the training loop, preserving data confidentiality without sacrificing detection performance.
• Adversarial robustness: empirical evidence that quantum‑based representations are harder to fool than conventional ML models.
• Comprehensive evaluation: benchmarks on multiple smart‑grid anomaly datasets (e.g., fault injection, cyber‑attack simulations) showing superior precision/recall and lower false‑positive rates.
• Scalability analysis: demonstrates that partitioning reduces circuit depth and qubit requirements, making the approach viable for realistic grid sizes.

Methodology

1. Data preprocessing – Raw sensor streams (voltage, current, frequency, etc.) are normalized and windowed into fixed‑size feature vectors.
2. Quantum feature map – Each vector is encoded into a quantum state using a parameterized feature‑embedding circuit (e.g., angle encoding with Pauli‑X rotations).
3. Partitioned network – The full quantum model is divided into k smaller sub‑circuits (partitions). Each partition processes a slice of the feature vector, and their outputs are classically aggregated (concatenation → fully‑connected layer). This reduces the number of qubits needed per run and shortens circuit depth, mitigating noise on NISQ devices.
4. Training loop – A hybrid quantum‑classical optimizer (e.g., Adam with parameter‑shift gradients) updates both quantum gate parameters and classical read‑out weights.
5. R‑QUPID (DP variant) – Gaussian noise is added to the gradient estimates in accordance with the moments accountant, guaranteeing a formal (ε,δ) differential‑privacy budget.
6. Evaluation – Models are compared against classical baselines (Random Forest, LSTM, XGBoost) on detection metrics (AUROC, F1‑score) and robustness tests (FGSM/PGD adversarial attacks, privacy leakage).

Results & Findings

• Higher detection quality: QUPID consistently outperforms the best classical baseline by 4–5 % AUROC and reduces false alarms by ~30 %.
• Robustness to attacks: Under FGSM with ε=0.1, QUPID’s AUROC drops only 1.2 % vs. >6 % for the LSTM.
• Privacy‑preserving performance: Adding DP (ε=1) incurs a negligible loss (<1 % AUROC), confirming that the quantum representation can absorb noise without collapsing.
• Scalability: Partitioning from 8‑qubit to 4‑qubit sub‑circuits cuts circuit depth by ~45 % and enables training on IBM’s 27‑qubit device with <5 % error rates.

Practical Implications

• Real‑time grid monitoring: Operators can deploy QUPID on edge quantum processors (or cloud‑based quantum services) to flag faults or cyber intrusions faster than classical pipelines that require heavy feature engineering.
• Reduced false alarms: Lower false‑positive rates translate directly into fewer unnecessary shutdowns and maintenance trips, saving operational costs.
• Security‑by‑design: The inherent adversarial resistance means that attackers would need to craft far more sophisticated perturbations, raising the bar for grid‑targeted cyber‑attacks.
• Privacy compliance: R‑QUPID satisfies differential‑privacy guarantees, helping utilities meet regulatory requirements (e.g., GDPR‑style data protection) while still leveraging rich sensor data.
• Scalable quantum deployment: The partitioned approach sidesteps the qubit‑count bottleneck, making it feasible to integrate quantum inference into existing SCADA/EMS stacks without waiting for fault‑tolerant quantum computers.

Limitations & Future Work

• Hardware constraints: Experiments rely on simulated noise models and modest‑scale NISQ devices; performance on larger, noisier quantum hardware remains to be validated.
• Partitioning overhead: While depth is reduced, the need for classical aggregation introduces latency that may be non‑trivial for ultra‑low‑latency protection schemes.
• Domain generalization: The study focuses on a handful of smart‑grid datasets; broader validation across different grid topologies and international standards is needed.
• Future directions: The authors propose exploring adaptive partitioning (dynamic allocation of qubits per feature), hybrid quantum‑classical ensembles, and extending the framework to other cyber‑physical domains such as water‑distribution or autonomous transportation networks.

Authors

• Hoang M. Ngo
• Tre’ R. Jeter
• Jung Taek Seo
• My T. Thai

Paper Information

• arXiv ID: 2601.11500v1
• Categories: cs.LG
• Published: January 16, 2026
• PDF: Download PDF