AAISM: Advanced in AI Security Management

A distillation attack uses synthetic examples and target-model outputs to transfer the target model’s behavior into a replica or student model. Model inversion instead seeks to reconstruct private training data or features, while poisoning modifies training data and prompt attacks manipulate inputs.