Deep reinforcement learning(DRL) has become a hotspot in the field of multi-robot systems due to its efficient performance. However, when encountering unstructured environment with time-varying and unknown risks, the traditional DRL methods exposes the disadvantage of poor risk defense ability and fragile system security. The unknown risk will bring nonlinear intrusion to the state space of multi-robot systems in the form of anti attack, which will pose a serious threat to the estimation of robot motion strategy. To solve this problem, this paper proposes a multi-agent reinforcement learning method based on active risk defense mechanism(ARD-MARL). Firstly, based on the locally observable Markov game model, a risk discrimination mechanism with global communication information is established to predict the current behavior state. Secondly, in the strategy deployment stage, we build an event-triggered multi risk processing scheme to implement the matching security strategy for different levels of risk prediction. Then, aiming at the dangerous state with risk intrusion, an active defense Actor-Critic network architecture based on the enhanced attention mechanism is designed. Through magnifying the important information and restraining the threat information, a safer and more efficient motion strategy is generated. Finally, extensive experiments are carried out in multi-agent cooperative and confrontation tasks. The results show that the multi-robot reinforcement learning method with active security defense mechanism can effectively enhance the stability and anti risk ability, and improve the security of information transmissions.