Abstract

A SURVEY OF DEEP LEARNING ARCHITECTURES FOR TRANSACTION AND FINANCIAL FRAUD DETECTION, ANCHORED ON THE MULTI-TASK CNN BEHAVIOURAL EMBEDDING MODEL

Ms. S. Jayanthi*, Dr. S. Jeyalaksshmi

ABSTRACT

This survey reviews deep learning architectures for transaction and financial fraud detection published strictly after 2024, using the Multi-task CNN Behavioural Embedding Model (MTCNN) proposed by Qu et al.[1] in 2024 as a conceptual anchor rather than as one of the post-2024 studies under review. MTCNN's core ideas — multi-range convolutional kernels, positional encoding, and multitask learning via random loss weighting, validated at production scale — are used as a lens through which 14 papers published in 2025 and 2026 are organized and compared. These recent works cluster into four architectural families: Transformer-based models (including a production-validated multi-stream fusion Transformer), Graph Neural Network models that expose relational fraud patterns invisible to purely sequential architectures, a lightweight dilated Temporal Convolutional Network (TCN) with built-in explainability, and hybrid CNN/RNN/ensemble models paired with explainable AI (XAI) tooling. We present two comparative tables — one organized by architectural family and one listing all 15 surveyed papers (the MTCNN anchor plus 14 post-2024 studies) individually — and discuss datasets, evaluation protocols, open challenges, and future directions, including the largely unexplored question of whether MTCNN's production-validated multitask CNN philosophy can be combined with the graph- and Transformer-based relational modelling that now dominates the post-2024 literature.

Keywords: Transaction Fraud Detection, Financial Fraud, Multi-task Learning, Convolutional Neural Networks, Transformers, Graph Neural Networks, Temporal Convolutional Networks, Explainable AI, Survey.