About Neural systems vulnerabilities: Classical attacks and recent defenses.

Abstract: The widespread adoption of voice-controlled devices and speech recognition systems underscores the critical need for robust security measures to ensure their reliability. These systems face growing threats from adversarial and poisoning attacks, which exploit vulnerabilities to degrade performance or manipulate outcomes. This talk explores the evolving landscape of adversarial attacks on speech systems, focusing on their detection and classification to illuminate their characteristics and impacts. We also investigate dirty and clean label poisoning attacks, where malicious data is stealthily introduced into training datasets, compromising system integrity. Finally, we present a range of defense mechanisms designed to counteract poisoning attacks, enhancing the resilience and trustworthiness of speech recognition technologies.