New framework TransferBreaker combats adversarial attacks on speech recognition models

Researchers have developed a new framework called TransferBreaker to enhance the security of fine-tuned Automatic Speech Recognition (ASR) models. These models, often deployed in black-box settings, are vulnerable to adversarial attacks where perturbations crafted for a base model can significantly degrade the performance of a fine-tuned version. TransferBreaker integrates several techniques, including Base Adversarial Fine-Tuning and Latent Jacobian Regularization, to suppress this adversarial transferability. Evaluations across multiple languages and ASR models demonstrated a substantial reduction in word error rate under adversarial conditions. AI

IMPACT Enhances the security and reliability of deployed speech recognition systems against adversarial manipulation.

RANK_REASON The cluster contains a research paper detailing a new method for improving model robustness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Base Adversarial Fine-Tuning
• Hugging Face
• HybridGrad-AFT
• Latent Jacobian Regularization
• TransferBreaker

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →