Learning Apparatus Mitigating Membership Inference Attacks via Training Data Modification

The patent applies preliminary action by modifying the training data before the distillation learning process begins. Specifically, it generates artificial training data with controlled target variable distributions that prevent overfitting from occurring during the learning process. This preliminary data preparation ensures that even during intermediate learning stages, the model remains resistant to membership inference attacks, rather than only achieving this property after complete training.

The patent implements preliminary anti-action by introducing a countermeasure against membership inference attacks before the harmful effect can manifest. By modifying the target variable distribution in advance during data generation, the system prevents the model from developing vulnerabilities to membership inference attacks during the learning process, rather than attempting to fix the problem after overfitting occurs.

The patent applies parameter changes by modifying the distribution parameters of the target variable during training data generation. Instead of using the original target variable distribution from the pre-learned model, it generates a new distribution with adjusted parameters that prevent the model from mistakenly learning in the direction of the original training data targets. This allows distillation to proceed effectively while maintaining correct learning direction.

The patent introduces an intermediary element - artificially generated training data with controlled target variable distributions - that mediates between the pre-learned model's output and the final distilled model. This intermediary data layer allows the distillation process to compress the model while preventing direct overfitting to the original training data targets, as the intermediary data provides a controlled learning signal.