In surveillance systems, pedestrians are often occluded by various obstacles. To address this issue, an occluded pedestrian re-identification method based on an adversarial defense network was proposed, which regards occlusion as noise interference and solves the problem from the perspective of adversarial defense. First, three methods: random erasure interference, random transformation interference, and random noise interference were introduced to generate three types of adversarial feature information with incomplete details, protecting the pedestrian re-identification system from the interference of information loss. Second, a multi-level attention mechanism was put forward. This module compensates for the information loss caused by occlusion by paying more attention to the occluded body parts or utilizing contextual information, thereby improving the recognition performance of the model in occluded scenarios. Finally, through the image high-low frequency feature enhancement module, the high-low frequency feature information of occluded pedestrians was integrated and strengthened, further boosting the model's recognition performance. Ablation experiments on the proposed method and experimental results on relevant datasets prove its effectiveness.