GAN-Based Training Data Recovery for Pre-Trained Classifiers

The patent creates a copy of the training data distribution by training a GAN to generate synthetic data samples that follow the same distribution as the original training data. The generator model learns to produce data samples that are statistically indistinguishable from the original training data, enabling quality evaluation without accessing the original data.

The patent introduces a GAN-based intermediary system that acts as a mediator between the deployed classifier model and the need for training data evaluation. The GAN generates synthetic training data that serves as a proxy for the original training data, allowing quality assessment of the classifier without direct access to the original training dataset.

The patent inverts the typical GAN training approach by using the pre-trained classifier model as part of the training objective. Instead of training the GAN to only match the training data distribution, the classifier loss function guides the generator to produce data that the classifier would have seen during its training, enabling recovery of unseen training data characteristics.

The patent implements a feedback mechanism where the pre-trained classifier model provides classification signals back to the GAN generator during training. The classifier loss computed on generated samples feeds back to refine the generator, creating a closed-loop system that enables the GAN to learn and reproduce the training data distribution even for classes not explicitly present in the generation process.