<p>With increasing concerns over privacy protection, existing methods face significant challenges in securing images containing multiple faces. This paper addresses this issue by proposing a multiple-face encryption framework that integrates precise face localization with a robust encryption scheme. For localization, an enhanced YOLOv8 model is developed, where the EMA attention mechanism captures fine-grained features, the BiFPN module enables efficient multi-scale fusion, and the WIoU loss reduces training errors. These improvements substantially enhance detection performance on face datasets, ensuring reliable identification of multiple facial regions. For encryption, a novel superposed algorithm based on a non-adjacent logistic-dynamic coupled map lattice is introduced, effectively overcoming the limitations of the Arnold cat map in handling rectangular images. Moreover, a two-way diffusion mechanism is incorporated to resist chosen-plaintext attacks, thereby enhancing security. Experimental results demonstrate that the proposed method achieves strong robustness against various cryptographic attacks, while providing a promising direction for advancing research in face image cryptography.</p>