Detecting Phishing in Ethereum Networks using Quantum Machine Learning

Research Question and Approach

This paper investigates the application of Quantum Machine Learning (QML) to the critical problem of phishing detection in Ethereum blockchain networks. Given the limitations of current quantum hardware, the authors evaluate two primary QML models: the Variational Quantum Classifier (VQC) and the Quantum Support Vector Machine (QSVM). To address the challenge of data encoding, they propose a novel 'cascaded' Quantum Random Access Coding (QRAC) scheme, which maps classical features into quantum states more efficiently than traditional methods like the ZZ feature map. The researchers benchmark these models against classical state-of-the-art techniques, such as Graph Convolutional Networks (GCN), and propose a stacking-based hybrid ensemble that combines the strengths of both quantum and classical algorithms.

Methodology

The authors utilize a dataset of 3 million Ethereum nodes, extracting seven statistical features (e.g., in-degree, out-strength, and total neighbors) to identify phishing activity. Because the dataset is heavily imbalanced, they construct a balanced training set of 320 samples to train their models. The study employs a two-level stacking ensemble: base models (QSVM and GCN) generate predictions, which are then used as input features for a meta-classifier. Hardware validation was performed on the IBM Heron quantum processor, using Qiskit Runtime primitives to manage circuit execution and minimize overhead. The cascaded QRAC encoding was specifically designed to reduce qubit requirements while maintaining feature distinguishability by distributing features across orthogonal planes on the Bloch sphere.

Key Findings

The results demonstrate that the QSVM consistently achieves the lowest false negative rates, making it particularly effective for anomaly detection where missing a phishing event is costly. The proposed cascaded QRAC encoding consistently outperformed the standard ZZ feature map, yielding accuracy improvements of 13% for VQC and 3% for QSVM. Furthermore, the authors observed that certain QML algorithms exhibit remarkable resilience on the IBM Heron processor, with performance metrics approaching those of noiseless simulators, suggesting that QML is increasingly viable for practical financial security tasks despite current hardware constraints.