Breaking Adversarial Transferability in Fine-Tuned Speech Recognition
TransferBreaker is a framework designed to stop adversarial attacks from transferring from a base speech recognition model to a fine-tuned target model. It uses three methods—Base-AFT, Latent Jacobian Regularization, and HybridGrad-AFT—to create robust defenses. The results show it drastically reduces performance drops under adversarial conditions across multiple languages and models.