LightFraud a lightweight hybrid deep learning model for real time fraud detection on edge devices

Abstract Fraud and anomaly detection have become increasingly important across financial, cybersecurity and healthcare domains due to the rapid growth of digital services and interconnected systems. Although deep learning-based approaches have demonstrated strong detection performance, their deployment in resource-constrained environments such as Internet of Things (IoT) devices, mobile platforms, and edge computing systems remains challenging due to high computational and memory overhead. This paper presents LightFraud, a lightweight hybrid deep learning framework designed for real-time fraud detection under strict resource constraints. LightFraud combines CNN-based local feature extraction with LSTM-based sequential modelling, along with dimensionality reduction and implicit graph-based regularization to improve learning stability without relying on explicit relational graph construction. To support edge deployment, the framework incorporates model compression techniques, including pruning and quantization to reduce memory footprint and inference latency. The proposed framework is evaluated across three heterogeneous domains, namely financial transactions, network intrusion detection, and healthcare event monitoring, using publicly available datasets. Experimental results show that LightFraud achieves competitive detection performance while reducing computational complexity and deployment overhead compared to conventional deep learning models. Deployment experiments on Raspberry Pi and NVIDIA Jetson Nano further demonstrate the feasibility of LightFraud for real-time edge-based fraud detection.