Deep Generative Model Watermarking via Hyperparameter Embedding

The patent applies parameter changes by modifying hyperparameters of the deep generative model to embed watermarks. Specifically, the watermark is embedded by adjusting parameters such as learning rates, weight decay, or other optimization hyperparameters during the training process. This allows the watermark to be incorporated into the model's operational characteristics without fundamentally changing the model architecture, thereby protecting intellectual property while managing complexity through parameter-level modifications rather than structural changes.

The patent applies preliminary action by embedding the watermark during the model training process rather than adding it afterward. The watermarking is performed as part of the training routine, where the watermark signal is incorporated into the training data or training process itself. This preliminary integration ensures that the watermark is established before the model is deployed, avoiding the need for additional computational resources during inference and reducing overall computational burden.

The patent applies the intermediary principle by introducing a watermark embedding mechanism that acts as a mediator between the training process and the final model. This intermediary component allows the watermark to be embedded through modifications to the training process itself (such as adjusting hyperparameters or incorporating watermark signals during training) rather than directly modifying the training dataset. The intermediary approach preserves model utility by maintaining the original training data integrity while achieving watermark embedding through the training process modifications.