The intersection of artificial intelligence, quantum computing and cybersecurity is a space ripe for innovation, and the recent hackathon sponsored by NovaceneAI and hosted on the Aqora platform showcased this potential. Aqora’s mission is to apply quantum computing to real-world challenges, providing a platform where participants can develop practical solutions, showcase their expertise, and collaborate with global experts in the field.
Quantum computing leverages the principles of superposition and entanglement to perform complex computations far more efficiently than classical systems. This advantage is particularly useful for machine learning applications, where quantum algorithms can optimize computations and uncover patterns in vast datasets. NovaceneAI presented participants with a real-world cybersecurity challenge: develop a quantum binary classifier model to detect suspicious network activity, leveraging the open-source BETH dataset from Highnam et al. As security alerts continue to rise in volume, organizations struggle to add enough skilled analysts to keep pace. NovaceneAI helps clients address this challenge by providing predictive systems that enhance efficiency and help teams focus on the most pressing security threats.
The Challenge: Detecting Suspicious Network Activity
Network security is an ongoing arms race between defenders and malicious actors. The ability to accurately classify suspicious activity is crucial for organizations looking to safeguard their systems. The BETH dataset, a rich source of network telemetry data, provided an excellent foundation for competitors to tackle this real-world issue. Participants were tasked with building a quantum binary classifier to distinguish between normal and suspicious network activity. Given the complexity of the data, the challenge required a blend of quantum computing techniques, domain knowledge, feature selection strategies, and optimization methods to maximize classification accuracy.
The Competitors and Their Innovative Approaches
Two competitors rose to the occasion and delivered outstanding solutions, ultimately tying for first place. Both participants built highly effective classifiers, using different feature engineering strategies.
Oleksii Adamov took a quantum one-class SVM approach.
- Dataset & Preprocessing: Tackled the BETH cybersecurity dataset (~763,000 login events) in an anomaly detection setting. The training data contained only normal logins (no attack instances), requiring a one-class classification strategy. Oleksii applied extensive preprocessing, combining feature extraction techniques from the original BETH paper with custom feature engineering to refine input variables for the model.
- Quantum Model: Implemented a quantum one-class Support Vector Machine (SVM) to learn the profile of “normal” login behavior and flag deviations. This approach leverages a quantum kernel-based classifier, which maps data into a high-dimensional Hilbert space where anomalies can be separated from normal data. He used an 8-qubit ZZ Feature Map (with two layers of linear entanglement) to encode the engineered features as quantum states. Each classical feature vector was binary-encoded into qubit rotations, and the entangled feature map circuit captured complex feature interactions in the quantum kernel.
- Kernel Optimization: To handle the large dataset efficiently, Oleksii employed a Nyström approximation to estimate the quantum kernel matrix in a hybrid quantum-classical manner. Instead of computing all pairwise kernel entries (which would be computationally infeasible for 763k samples), a subset of data was used to approximate the full kernel space. This sped up the kernel computation and made it feasible to train the one-class SVM on quantum features.
- Toolchain: The solution was built with Qiskit, utilizing 8 qubits in simulation. Quantum circuits were used to compute kernel values (overlap of statevectors) for training the SVM, and the resulting model could classify new logins by evaluating them with the learned quantum kernel. The pipeline demonstrated how quantum kernel methods can be applied to real cybersecurity data using current quantum software tools.
Peter Yang took a quantum k-means clustering approach.
- Preprocessing: Peter’s approach took a clustering angle to anomaly detection. He first applied Principal Component Analysis (PCA) on the BETH dataset to reduce dimensionality, isolating the features most correlated with login behavior. By retaining only the top principal components (the highest-variance dimensions), he obtained a compact data representation suitable for a small quantum register (around 4–5 qubits) while preserving key information about normal vs. abnormal patterns.
- Quantum Algorithm: Developed a hybrid quantum K-means clustering algorithm to group login events and detect outliers. Each data point (after PCA) was embedded as a quantum state across 4–5 qubits (e.g., using angle or amplitude encoding for each principal component). He initialized a set of $k$ cluster centroids also as quantum states. The core quantum routine used a swap test to evaluate distances: by entangling a data state with a centroid state and performing the swap test, the algorithm estimates the Euclidean distance between the two quantum state. This swap-test subroutine yields the inner product between state vectors as a similarity measure, which is used to determine the closest centroid for each data.
- Iterative Clustering: The quantum distance evaluations were integrated into the classic K-means iterative loop. After assigning each quantum-encoded data point to its nearest centroid (based on swap-test distance), the cluster centroids were updated classically (recomputed as the mean of all points in the cluster, then re-prepared as new quantum states). These steps repeated until convergence, resulting in clusters of normal logins and identification of anomalous points as those not fitting any cluster well. Peter’s method thus combined quantum distance computation with classical centroid updates, illustrating a cooperative quantum-classical workflow.
- Toolchain: The solution was prototyped on a local simulator (using ~4–5 qubits, e.g. with Qiskit) rather than on cloud quantum hardware, due to the experimental nature of the algorithm. The quantum swap-test circuits for distance measurement were run in simulation, and their outputs fed into the classical K-means procedure. This hybrid implementation allowed testing quantum subroutines for clustering on realistic data, showing how quantum computing could be applied to anomaly detection via unsupervised learning with current qubit limits.
Key Takeaways and Future Directions
The tie for first place highlights the diversity of solutions possible in quantum machine learning-based cybersecurity. While traditional feature engineering offers explainability and control, automated techniques present an avenue for discovering hidden patterns that may otherwise go unnoticed. Future advancements in quantum computing and AI may further enhance these capabilities, paving the way for more robust and adaptive security models. A presentation detailing the solutions and insights from the challenge is available here: Watch Now.
NovaceneAI and Aqora’s collaboration underscores the importance of community-driven innovation in advancing quantum computing, AI and cybersecurity. As technology continues to evolve, hackathons like this will remain invaluable in pushing the boundaries of what’s possible.
Stay tuned for future challenges where quantum and AI converge to solve some of the world’s toughest problems!
