Privacy-Preserving Semantic Communication via Adversarial Discriminators

The patent introduces an adversarial discriminator as an intermediary component that prevents direct privacy attacks on the transmitted features. The discriminator acts as a mediator between the transmitter and potential adversaries, learning to distinguish real transmitted features from generated ones, thereby protecting privacy without directly degrading the semantic communication performance.

The patent changes the training parameters and objective functions of the semantic communication system by incorporating adversarial loss terms. Instead of using traditional privacy protection parameters like differential privacy epsilon, the system uses adversarial game theory parameters (generator loss, discriminator loss) to achieve privacy protection while maintaining performance.

The patent applies preliminary anti-action by training an adversarial discriminator before deployment to learn the distribution of legitimate transmitted features. This pre-trained discriminator serves as a defense mechanism that prevents model inversion attacks by making it computationally difficult for adversaries to reconstruct original inputs from transmitted features.

The patent converts the potential harm of transmitted features being used for privacy attacks into a benefit by using adversarial training. The same transmitted features that could be exploited by adversaries are now used to train the discriminator, which then protects the system by identifying and preventing such attacks, turning the vulnerability into a security feature.