# A deep residual 1D-CNN with self-attention for fraud transaction detection in virtual economies

## TL;DR

This paper introduces a deep learning model for detecting fraud in metaverse transactions using a 1D-CNN with self-attention, achieving strong performance on benchmark datasets.

## Contribution

A novel 1D-CNN with residual connections and self-attention for real-time fraud detection in metaverse financial transactions.

## Key findings

- The model achieved excellent accuracy, sensitivity, and specificity on metaverse financial datasets.
- It maintained strong performance on the Credit Card Fraud Detection dataset.
- Ablation studies confirmed its robustness to noisy data and real-world imperfections.

## Abstract

As virtual economies in the metaverse continue to grow, the need for real-time risk assessment in financial transactions has become critical. Traditional fraud detection systems often face challenges in keeping pace with the complexity and speed of metaverse data. To address this, we introduce a real-time anomaly detection and risk classification model designed specifically for metaverse transactions. The model is based on a one-dimensional convolutional neural network (1D-CNN) enhanced with residual connections and a self-attention mechanism, allowing it to focus on the most relevant features of each transaction for improved risk classification. We trained the model on benchmark metaverse financial datasets from Kaggle, achieving excellent results in accuracy, sensitivity, and specificity when classifying transactions into three risk levels—low, moderate, and high. To validate its robustness, we also tested it on the widely used Credit Card Fraud Detection dataset, where it maintained strong performance. However, we acknowledge that perfect scores can sometimes indicate overly clean or predictable data. To address this, we conducted an ablation study by introducing controlled noise into the dataset, evaluating the model’s ability to handle uncertainty and imperfections in real-world scenarios. To enhance interpretability, we analyzed feature importance across several CNN-based variations and assessed performance using confusion matrices, ROC curves, and t-SNE visualizations, which confirmed clear separation of risk levels in high-dimensional space. Further comparisons with other machine learning and deep learning models demonstrate the confidence and effectiveness of the proposed 1D-CNN architecture for financial fraud detection in the metaverse.