Multi-Scale 1D Convolutional Networks for Robust Detection of Anomalous Financial Transactions

Abstract

This study addresses the challenge of identifying fraudulent behavior in financial transactions by proposing a time series modeling method based on one-dimensional convolutional neural networks (1DCNN) for efficient detection of complex and concealed anomalous patterns. The method divides raw transaction sequences into local subsequences using a sliding window mechanism to construct a time-aware input structure, guiding the model to learn latent behavioral differences within local contexts. A multi-scale convolutional architecture is incorporated to extract short-term variations and mid-range behavioral patterns in parallel using different receptive fields, enhancing the modeling of temporal dependencies. After feature extraction, the model applies flattening and fully connected layers for nonlinear mapping and uses a weighted binary cross-entropy loss function to optimize classification performance under imbalanced class distributions. The study constructs multiple experimental dimensions, including sliding window length variation, multi-scale structure configuration, and robustness analysis under class imbalance, to comprehensively evaluate model performance in financial fraud detection scenarios. The results show that the proposed method, without relying on prior rules or complex feature engineering, can extract effective representations from transaction sequences and demonstrates strong capability in detecting abnormal behavior, highlighting the modeling advantages and applicability of deep neural networks in financial risk control tasks.