Information Processing Device for Real-Space Attack Sample Generation

The patent introduces an autoencoder as an intermediary model that learns to map high-dimensional and nonlinear feature vectors to a lower-dimensional latent space. This intermediary representation simplifies the feature space, making it easier to generate attack samples by modifying latent vectors and then decoding them back to the original feature space, thereby improving efficiency without losing the ability to represent complex attack patterns.

The patent applies preliminary action by using the autoencoder to pre-process and transform real attack data into synthetic attack samples. The autoencoder learns the underlying patterns and distributions of attack data during training, then generates additional synthetic attack samples that preserve these patterns. This preliminary transformation enables the system to create sufficient training data without requiring large quantities of actual attack data.

The patent uses copying by generating synthetic attack samples that replicate the patterns and characteristics of real attack data. The autoencoder learns from a small set of real attack samples and creates multiple synthetic copies that preserve the essential features of actual attacks. These synthetic copies provide diverse training examples without requiring increasingly sophisticated real attack data, thereby improving detection accuracy while maintaining a manageable level of attack sophistication.