Convolutional Neural Networks (CNNs) are widely used across various fields, but their parameters are susceptible to be tampered, threatening model security. In this paper, we propose a hierarchical recovery method for CNNs based on self-embedding watermarking to protect the model integrity. The method introduces an entropy-based parameter importance ranking mechanism and a differentiated blocking strategy: fine block division of important parameters improves recovery accuracy, and coarse block division of less important parameters reduces computational overhead. Authentication bits and reference bits are embedded into the Least Significant Bits (LSBs) of parameter blocks to enable efficient tampering detection and localization, and the damaged parameters are restored using untampered blocks. Experiments show that the proposed method achieves high detection accuracy and excellent recovery performance, verifying its effectiveness and reliability.