The United States Patent and Trademark Office (USPTO) has granted patent US12585915B2, titled 'Neural network training method, apparatus, device, storage medium,' to BEIJING BAIDU NETCOM SCIENCE TECHNOLOGY CO., LTD. The patent describes a method for training neural network models that involves acquiring encrypted feature representations and tag data, calculating loss error and gradient ciphertexts, and updating network parameters based on decrypted gradient information. This approach aims to enhance privacy and security during the training process.

This patent grant represents a new intellectual property right for a specific AI training methodology. While not a regulatory rule imposing obligations on other entities, it signifies a patented technology in the field of AI and data privacy. Companies developing or utilizing similar neural network training techniques, particularly those involving encrypted data or secure multi-party computation, should be aware of this patent to assess potential infringement risks. No immediate compliance actions are required for entities not directly involved in implementing this specific patented method, but it highlights the evolving landscape of AI innovation and associated IP considerations.