With the acceleration of healthcare information digitalization, medical imaging data have increased dramatically. As these data contain sensitive patient information, ensuring their security during storage and transmission has become critically important. To address the privacy protection challenges of medical images, this paper proposes a medical image zero-watermarking algorithm based on the efficient convolutional neural network EfficientNet-B0. First, the EfficientNet-B0 model is fine-tuned by replacing the original 1000-dimensional fully connected layer, Softmax layer, and classification layer with a 128-dimensional fully connected layer and a regression layer. Next, the fine-tuned EfficientNet-B0 model is trained on the medical image dataset in this study to extract feature vectors of medical images. Then, the all-phase discrete cosine biorthogonal transform (APDCBT) is applied to the feature vectors and combined with a per-ceptual hashing algorithm to generate robust feature matrices. Finally, the original watermark image is encrypted using the chaotic sine-logistic-tent system (CSLTS), and the encrypted watermark is then XORed with the robust feature matrix to obtain the final authentication zero-watermark. The experimental results show that the proposed algorithm achieves strong robustness against common attacks with different intensities. Furthermore, the proposed algorithm outperforms other comparison algorithms in overall robustness in the comparative experiments.