VAE Training with Latent Regularization Against Adversarial Noise

The patent applies preliminary action by adding noise to the training data before training the VAE. This preprocessing step creates noisy training data that forces the model to learn robust latent representations from the outset, rather than attempting to defend against adversarial inputs after training. The noise addition is performed in advance as part of the training pipeline, embedding robustness directly into the learning process.

The patent modifies the training parameters by introducing a regularization term that explicitly controls the distance between latent representations of original and noisy data pairs. This parameter change transforms the objective function to penalize large deviations in latent space, directly optimizing for robustness. The regularization strength can be tuned to balance robustness gains against training complexity.

The patent applies local quality by differentiating the treatment of different data aspects during training. The loss function combines two distinct components: one that prioritizes maintaining stable latent representations (for classification accuracy) and another that handles reconstruction. By applying localized regularization specifically to the latent space while maintaining separate reconstruction objectives, the model can optimize classification performance without uniformly compromising reconstruction quality across all dimensions.

The patent uses partial action by applying regularization selectively to reduce latent representation distance for certain types of noise or data pairs, rather than uniformly constraining all reconstructions. This partial regularization approach allows the model to maintain classification accuracy where it matters most while preserving reconstruction capability for other aspects of the data, avoiding excessive constraints that would degrade overall performance.