Design of an E-commerce Malicious Behavior Recognition and Active Defense System Based on Multimodal Deep Learning and Dynamic Adversarial Game

Abstract

The rapid expansion of e-commerce platforms has been accompanied by increasingly sophisticated malicious attacks, including fraudulent transactions, account takeovers, and algorithmic manipulation, necessitating advanced defensive measures. This paper presents the design of a comprehensive identification and defense system for e-commerce malicious attacks based on deep learning architectures. The proposed system employs a multi-modal neural network framework that simultaneously processes heterogeneous data streams—user behavior sequences, transaction patterns, and real-time platform interactions—to detect anomalous activities with high precision. Specifically, we implement a hybrid model combining Long Short-Term Memory (LSTM) networks for analyzing temporal dynamics in user behavior, Graph Neural Networks (GNNs) for identifying suspicious relational structures among accounts, and Transformer-based encoders for detecting subtle semantic manipulations in review and listing content. The system is further enhanced by an online learning module that continuously adapts to emerging attack strategies, ensuring sustained robustness in dynamic threat environments. In validation experiments conducted on a large-scale e-commerce dataset comprising over 5 million user events, our approach achieved a recall of 94.2% and a false positive rate of just 1.8%, significantly outperforming conventional rule-based and standalone machine learning methods. The integrated defense mechanism not only identifies malicious entities but also triggers automated countermeasures such as session termination, challenge-based authentication, and behavioral quarantine. This work establishes a scalable, intelligent security framework suitable for modern e-commerce ecosystems and contributes a viable paradigm for deploying deep learning in business-critical protection systems.