Fraud Detection on Ethereum Transaction Using Hybrid Oversampling with Graph Neural Network

Sabrina, Iffa Amalia (2026) Fraud Detection on Ethereum Transaction Using Hybrid Oversampling with Graph Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221077-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Blockchain is a distributed ledger technology where transactions are stored in a structure of block linked chronologically through cryptographic hashes, ensuring data integrity and immutability. Ethereum, as one of the most prominent blockchain platforms, supports both Ether-based cryptocurrency transactions and smart contracts, making it a frequent target for sophisticated fraud schemes. The rapid growth of Ethereum-based financial transactions has led to an increase in advanced cryptocurrency fraud, creating significant challenges for automated detection tools. A crucial or primary obstacle is severe class imbalance, with only 6.1% of addresses representing error-related transactions out of all records. This imbalance causes the performance of Graph Neural Networks (GNNs) to decline sharply. This research presents a hybrid oversampling framework that transforms 254,973 Ethereum transaction records into a directed weighted graph, where nodes represent addresses enriched with behavioural features (including transaction counts, amount statistics, and temporal patterns such as transaction velocity and burst activity) and edges represent transactions between addresses. Node labels are derived from transaction-level “isError” flags through address-level aggregation, producing a transparent and authentic proxy labels for error-related address dataset. The framework combines Graph-aware SMOTE with a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) using a per-node adaptive mixing ratio calculated from local density, prediction uncertainty, majority-class separation, neighbour fraud ratio, and node degree, rather than a single fixed mixing parameter. Three GNN architectures, which are Graph Convolutional Network (GCN), Graph Attention Network (GAT), and GraphSAGE, were evaluated under four conditions, which are Baseline, GraphSMOTE only, WGAN-GP only, and Hybrid, across three random seeds (42, 123, 2024) on the Ethereum transaction dataset. F1-score and AUC-PR are adopted as the primary evaluation metrics in this research, chosen for their robustness under severe class imbalance. Accuracy is additionally reported for interpretability. Hybrid oversampling achieved the best mean F1-score of 0.5330 (GAT) and AUC-PR of 0.4902 (GAT), outperforming all other conditions, with the best mean or average AUC-ROC of 0.9382 (GAT). Temporal behavioural features improved the mean F1-score from 0.5085 to 0.5191 and mean AUC-PR from 0.4232 to 0.4735, indicating that temporal behavioural features provide useful discriminative signals for fraud detection. The proposed per-node adaptive mixing ratio yields a validation F1-score of 0.5842, compared to 0.5552 for the best fixed alpha (α=0.80), confirming that adaptive balancing of the SMOTE–WGAN-GP contributions according to local node characteristics outperforms a single global mixing ratio.
================================================================================================================================
Blockchain adalah teknologi buku besar terdistribusi (distributed ledger technology) di mana transaksi disimpan dalam struktur blok yang dihubungkan secara kronologis melalui hash kriptografi, memastikan integritas dan kekebalan data. Ethereum, sebagai salah satu platform blockchain yang paling menonjol, mendukung transaksi mata uang kripto berbasis Ether maupun smart contract, menjadikannya target yang sering digunakan untuk skema penipuan canggih. Pertumbuhan pesat transaksi keuangan berbasis Ethereum telah menyebabkan peningkatan penipuan mata uang kripto tingkat lanjut, menciptakan tantangan signifikan bagi alat deteksi otomatis. Hambatan utama atau krusial adalah ketidakseimbangan kelas (class imbalance) yang sangat tinggi, di mana hanya 6,1% alamat yang merepresentasikan transaksi terkait kesalahan (error-related transactions) dari seluruh data. Ketidakseimbangan ini menyebabkan kinerja Graph Neural Network (GNN) menurun secara signifikan. Penelitian ini menyajikan kerangka kerja oversampling hibrida yang mengubah 254.973 catatan transaksi Ethereum menjadi graf berbobot terarah, di mana simpul mewakili alamat yang diperkaya dengan fitur perilaku (termasuk jumlah transaksi, statistik nilai transaksi, dan pola temporal seperti kecepatan transaksi serta aktivitas lonjakan transaksi (burst activity)), sedangkan edge merepresentasikan transaksi antar alamat. Label node diperoleh dari penanda atau flag “isError” pada tingkat transaksi melalui agregasi pada tingkat alamat, sehingga menghasilkan label proksi yang transparan dan autentik untuk dataset alamat yang terkait kesalahan transaksi. Kerangka kerja ini menggabungkan Graph-aware SMOTE dengan Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) menggunakan rasio pencampuran adaptif per node (per-node adaptive mixing ratio) yang dihitung berdasarkan kepadatan kepadatan lokal, ketidakpastian prediksi (prediction uncertainty), pemisahan dari kelas mayoritas, rasio tetangga fraud, dan derajat node, bukan menggunakan satu parameter pencampuran tetap secara global. Tiga arsitektur GNN, yaitu Graph Convolutional Network (GCN), Graph Attention Network (GAT), dan GraphSAGE, dievaluasi dalam empat kondisi, yaitu Baseline, GraphSMOTE only, WGAN-GP only, dan Hybrid, menggunakan tiga random seed (42, 123, 2024) pada dataset transaksi Ethereum. Skor F1 (F1-score) dan AUC-PR digunakan sebagai metrik evaluasi utama yang digunakan dalam penelitian ini, dipilih karena ketahanannya (robustness) terhadap ketidakseimbangan kelas yang parah. Accuracy turut dilaporkan untuk kepentingan interpretasi. Oversampling hibrida mencapai skor F1 rata-rata terbaik sebesar 0,5330 (GAT) dan AUC-PR sebesar 0,4902 (GAT), mengungguli semua kondisi lainnya, dengan AUC-ROC rata-rata terbaik sebesar 0,9382 (GAT). Fitur perilaku temporal meningkatkan skor F1 rata-rata dari 0,5085 menjadi 0,5191 dan AUC-PR rata-rata dari 0,4232 menjadi 0,4735, menunjukkan bahwa fitur perilaku temporal memberikan sinyal diskriminatif yang berguna untuk deteksi penipuan. Rasio pencampuran adaptif per-node yang diusulkan menghasilkan skor F1 validasi sebesar 0,5842, dibandingkan dengan 0,5552 untuk alpha tetap terbaik (α=0,80), yang menegaskan bahwa penyeimbangan adaptif kontribusi SMOTE–WGAN-GP terhadap karakteristik node lokal mengungguli rasio pencampuran global tunggal.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Ethereum, Fraud Detection, Graph Neural Network, GraphSMOTE, WGAN-GP Deteksi Fraud, Ethereum, Graph Neural Network, GraphSMOTE, WGAN-GP
Subjects:	Q Science > QA Mathematics > QA275 Theory of errors. Least squares. Including statistical inference. Error analysis (Mathematics)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	IFFA AMALIA SABRINA
Date Deposited:	27 Jul 2026 08:12
Last Modified:	27 Jul 2026 08:12
URI:	http://repository.its.ac.id/id/eprint/137847

Actions (login required)

View Item