Artificial intelligence systems rely on large-scale, open training data collection, which creates opportunities for data poisoning attacks. Attackers can manipulate training data by infiltrating the data supply chain, thus compromising the availability of deep learning models. Compared to dirty-label data poisoning attacks based on label flipping, clean-label attacks are more stealthy but face greater challenges in implementation. On the one hand, precise manipulation of sample features is required rather than direct alteration of labels. On the other hand, it is necessary to ensure that poisoned samples remain highly visually consistent with the original samples to evade detection. Although existing clean-label attacks have achieved some success, they are often difficult to implement in practical scenarios due to loose assumptions, and there remains room for improvement in terms of stealthiness and computational efficiency. To address these issues, this study proposes a clean-label data poisoning attack scheme for deep learning based on data reconstruction, termed CPDR. In this scheme, the loss function of the Attention U-Net model, originally designed for image segmentation, is reconstructed and transformed into a multi-task learning reconstruction model that incorporates reconstruction loss, attack effectiveness loss, and model update consistency loss. Through data reconstruction, stealthy poisoned samples are generated, and effective poisoning attacks are achieved under a black-box assumption. Experiments conducted in both centralized and federated learning scenarios, across two datasets and four target models, demonstrate that CPDR reduces the model’s classification accuracy on target classes with minimal impact on the performance of the main task. In centralized learning, CPDR achieves an additional reduction of 2.33 and 2.42 percentage points in target class accuracy compared to VagueGAN and Adversarial Poison, respectively. In federated learning, this advantage further increases to 4.44 and 4.10 percentage points, respectively. Moreover, the poisoned samples generated by CPDR exhibit minimal visual differences from the original samples, making them difficult to distinguish and detect, thus achieving a good balance between stealthiness and attack effectiveness.