The current society is complex and changeable, and the post-pandemic era profoundly affects people's work and life. Identifying the potential risks of highrisk individuals in society and carrying out early warning and control work effectively is the focus of current public security work and is also the key to maintaining social stability and people's peace. This work first analyzes and constructs a knowledge graph of high-risk individuals based on their backgrounds, trajectories, and related information. Subsequently, we propose a high-risk personnel risk assessment model based on a graph attention-label propagation algorithm. The model employs a multi-label feature selection method, a basic classifier based on a graph attention network for the label propagation algorithm, and an adversarial data augmentation algorithm to enhance the gradient-based adversary during training. In the experiment, we train the model using a public-security-field personnel dataset, and the accuracy of the proposed method reaches 90.2%, Ablation experiments demonstrate the effectiveness and stability of the proposed method. Constructing a knowledge graph specifically for high-risk individuals based on backgrounds, trajectories, and related data,Proposing a risk assessment model using a graph attention-label propagation algorithm, incorporating multi-label feature selection and adversarial data augmentation, which enhances training effectiveness.