Implementing Large Quantum Boltzmann Machines as Generative AI Models for Dataset Balancing

The gist: This study explores the implementation of large Quantum Restricted Boltzmann Machines (QRBMs), a key advancement in Quantum Machine Learning (QML), as generative models on D-Wave’s Pegasus quantum hardware to address dataset imbalance in Intrusion Detection Systems (IDS).

Dataset Imbalance Challenge

Balanced datasets have long been recognized as essential for effective machine learning, enabling models to generalize and produce robust predictions. In the context of intrusion detection systems (IDS), the importance of balanced datasets becomes even more pronounced. Real-world datasets used for training IDS have often suffered from severe class imbalance, where the volume of normal traffic significantly outweighed that of intrusion attempts. This imbalance has limited the effectiveness of ML models, resulting in biased predictions favouring the majority class, reduced detection rates for rare but critical intrusions, and increased false positive rates. Traditional balancing methods such as oversampling and Synthetic Minority Oversampling Technique (SMOTE) have significant drawbacks. Oversampling introduced redundancy and increased the risk of overfitting; undersampling often discarded valuable information from the majority class. SMOTE can not handle class boundaries effectively and has high-dimensional or sequential data challenges.

QRBM Implementation on Quantum Hardware

The study successfully implemented a large QRBM (120 visible and 120 hidden units) on D-Wave’s Pegasus architecture, surpassing the limitations of default embedding tools. The QRBM synthesized over 1.6 million attack samples, achieving a balanced dataset of over 4.2 million records. The Pegasus topology provided enhanced qubit connectivity and increased computational capacity compared to earlier architectures like the Chimera topology. An algorithm was developed to implement large QRBMs (up to 172x120) on the Pegasus topology. This embedding algorithm outperforms D-Wave’s default embedding tool. The algorithm efficiently identifies multiple QRBM configurations (ranging from 4x4 to 172x120) in milliseconds.

QRBM as a Generative Model

RBMs can be used as generative models as they learn and replicate a dataset’s statistical properties. Once trained, an RBM can generate new data samples by alternately sampling from the visible and hidden layers in a process known as Gibbs sampling. This generative capability makes RBMs a powerful tool for several applications. Unlike oversampling techniques that merely duplicate existing data points, QRBMs generate entirely new samples that preserve the statistical diversity of the dataset.

Performance Evaluation and Comparison

The results showed significant improvements in detection rates and reductions in false positives when using QRBM-generated datasets. The QRBM achieved a balancing time of 0.33 seconds, which was faster than SMOTE (0.75 seconds) and Random Oversampler (0.23 seconds). When compared to classical methods, the QRBM produced higher-quality synthetic samples. For instance, in Table IX evaluation metrics after balancing using QRBM, the Random Forest achieved the highest precision at 96.18%, and the Decision Tree followed closely with a precision of 95.94%.

Conclusion on Quantum Advantage

QRBMs achieved superior performance compared to classical methods by generating high-quality synthetic samples that improved evaluation metrics like precision, recall, and F1 scores across multiple ML models. QRBMs leverage quantum hardware to model energy landscapes directly, offering rapid sampling and scalability. Future research should focus on overcoming limitations such as limited qubit connectivity and noise through advancements in quantum hardware. The introduction of the Zephyr topology has the potential to mitigate many of these hardware limitations by enhancing qubit connectivity and density. This study demonstrates the practical feasibility of large-scale QRBM implementations on existing quantum hardware. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. This work contributes by demonstrating the practical feasibility of large-scale QRBM implementations pushing the boundaries of what can be achieved on existing quantum hardware. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. The QRBMs achieved superior performance compared to classical methods by generating high-quality synthetic samples that improved evaluation metrics like precision, recall, and F1 scores across multiple ML models. The QRBMs exhibit clear advantages over classical methods such as SMOTE and RandomOverSampler by efficiently generating high-quality synthetic data using quantum annealing to model complex interactions. The QRBMs’ generative modelling capability, supported by the enhanced connectivity and flexibility of the Pegasus topology, enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. This study has demonstrated the potential of quantumrestricted Boltzmann Machines (QRBMs) as a powerful tool for addressing dataset imbalances e.g., in intrusion detection systems (IDS). The QRBMs generated high-quality synthetic data, significantly improving IDS performance across multiple metrics, including precision, recall, and F1 scores. The QRBMs provided superior results while maintaining remarkable computational efficiency. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. The QRBMs achieved superior performance compared to classical methods by generating high-quality synthetic samples that improved evaluation metrics like precision, recall, and F1 scores across multiple ML models. This study has demonstrated the potential of quantumrestricted Boltzmann Machines (QRBMs) as a powerful tool for addressing dataset imbalances e.g., in intrusion detection systems (IDS). The QRBMs generated high-quality synthetic data, significantly improving IDS performance across multiple metrics, including precision, recall, and F1 scores. The QRBMs provided superior results while maintaining remarkable computational efficiency. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. The QRBMs’ generative modelling capability enabled efficient sampling from complex energy landscapes, overcoming the limitations of classical approaches like Gibbs sampling. This study has demonstrated the potential of quantumrestricted Boltzmann Machines (QRBMs) as a powerful tool for addressing dataset imbalances e.g., in intrusion detection systems (IDS).

1. Improved data preprocessing pipeline using feature selection to reduce dimensionality before QRBM input. This step involves removing features with high correlations (absolute correlation greater than or equal to 0.9) and those with zero standard deviation, which helps in preparing the dataset for efficient embedding on quantum hardware as described in Section IV A.

2. Enhanced synthetic data generation quality by utilizing the QRBM's ability to synthesize new samples that preserve the statistical diversity of the dataset rather than just duplicating existing points. This capability ensures that generated data better captures intricate and highly non-uniform distributions, leading to superior precision, recall, and F1 scores compared to classical methods like SMOTE.

3. Faster dataset balancing time by leveraging the QRBM's rapid sampling capability of 1,000 samples in milliseconds per anneal. This efficiency allows the entire balancing task to be completed in just 0.33 seconds, significantly outperforming classical methods that took longer (e.g., SMOTE at 0.75 seconds).

4. More robust model training by using QRBM-generated datasets to train a wider variety of classifiers, including Support Vector Machines, Logistic Regression, and Gradient Boosting. The study shows that models trained on the QRBM-balanced data achieve significantly higher metrics (e.g., Random Forest precision of 96.18%) compared to those trained on unbalanced or SMOTE-balanced data.

5. Development of a scalable quantum embedding algorithm for larger Restricted Boltzmann Machines, specifically targeting architectures like the Zephyr topology mentioned in Section VI. This improvement would allow the implementation of larger and more complex QRBMs by optimizing chain lengths to remain short while maximizing qubit utilization, pushing the boundaries of what can be achieved on current quantum hardware.

This study explores the implementation of large Quantum Restricted Boltzmann Machines (QRBMs), a key advancement in Quantum Machine Learning (QML), as generative models on D-Wave's Pegasus quantum hardware to address dataset imbalance in Intrusion Detection Systems (IDS). By leveraging Pegasus's enhanced connectivity and computational capabilities, a QRBM with 120 visible and 120 hidden units was successfully embedded, surpassing the limitations of default embedding tools. The QRBM synthesized over 1.6 million attack samples, achieving a balanced dataset of over 4.2 million records. Comparative evaluations with traditional balancing methods, such as SMOTE and RandomOversampler, revealed that QRBMs produced higher-quality synthetic samples, significantly improving detection rates, precision, recall, and F1 score across diverse classifiers. The study underscores the scalability and efficiency of QRBMs, completing balancing tasks in milliseconds. These findings highlight the transformative potential of QML and QRBMs as next-generation tools in data preprocessing, offering robust solutions for complex computational challenges in modern information systems.