Storing palmprints in plaintext presents significant security risks due to their susceptibility to theft and misuse, yet research on secure storage methods remains insufficient. To address this gap, we propose a reversible encryption algorithm leveraging a controllable diffusion model. Our approach introduces a novel reversible texture transformation for palmprint images, supported by a new backbone architecture integrating self-attention mechanisms and SimAM. We further developed a transformation strategy that incorporates noise addition, pixel position swapping, and denoising to enhance encryption robustness and privacy preservation. The model guarantees reversibility, enabling secure decryption when necessary without compromising performance. Experimental results on the MPD palmprint dataset show that while the recognition accuracy of a model trained on original images for generated images is only 19.31%, a model trained on generated images achieves an accuracy of 88.44% for recognizing original images. Compared to existing methods, our approach demonstrates superior performance in both recognition accuracy and encryption effectiveness.