Prison management faces the problem of monitoring the behavior of incarcerated people, and the traditional monitoring methods have defects such as poor real-time performance and low confidentiality. This study proposes a behavioral monitoring method for prison inmates based on convolutional neural network image encryption technology. This method extracts the original features of the image, uses a filter for filter diffusion processing, combines convolutional neural network and compression perception theory to generate image hybrid phase mask, and finally realizes image encryption protection by replacing the image hybrid phase mask. The experiment adopts the self-constructed prison behavior dataset of inmates, which contains a total of 710 video samples of abnormal behaviors such as fighting, attacking the police, falling down and other daily behaviors. The results show that the proposed model has an accuracy of 94.14% in the assessment of the risk of assault, which is higher than 91.10% and 91.16% for the risk of suicide and the risk of negative rehabilitation. When the number of graph convolution layers is 2, the model performance is optimal with AUC value, accuracy, recall and F1 value of 96.26%, 92.17%, 88.25% and 85.72%, respectively. The analysis by manual intervention shows that the abnormal behaviors of prison inmates with appropriate intervention are significantly alleviated, while the non-interventionists still maintain the high sensitivity score status. The results of image encryption processing show that the pixel values of the ciphertext image are close to uniform distribution in the 0-255 interval, which effectively hides the statistical features of the plaintext image and ensures the security of the monitoring data. This study provides technical support for intelligent management and risk prevention and control in prisons, and has practical value for improving the modernized management level of prisons.
