Market malicious bidding user detection based on multi-agent reinforcement learning

With the rapid growth of e-commerce and online auction markets, malicious bidding activities have severely disrupted market order. Traditional detection methods face limitations due to their inability to effectively address the covert nature, dynamic characteristics, and massive data volumes associated with such behaviors. To address this challenge, this paper proposes a detection method for users engaging in malicious bidding based on Multi-Agent Reinforcement Learning (MARL). This approach first models target users as specialized agents, then integrates their historical bidding data, and finally learns optimal strategies through competitive games with adversarial agents. Additionally, this paper designs a dynamic adjustment mechanism for the maliciousness coefficient to simulate user behavior changes, enabling precise assessment of malicious intent. Compared to existing fraud detection approaches based on reinforcement learning, the fundamental innovation lies not merely in applying MARL technology, but in introducing the novel “dynamic maliciousness coefficient” mechanism. This mechanism achieves dynamic and precise maliciousness assessment through mathematical modeling and real-time iteration, addressing the shortcomings of traditional MARL models in capturing user behavioral heterogeneity. Experimental results demonstrate that this method exhibits higher detection accuracy and adaptability in complex dynamic market environments. It effectively captures bidder interaction relationships, significantly enhancing the detection of malicious behavior.