Voice processing systems, like automatic speech recognition (ASR) systems, have primary vulnerabilities that enable malicious voice attacks. These vulnerabilities include:
- Replay Attacks: Attackers record genuine human voice and replay it, bypassing authentication systems that rely on voice recognition.
- Spectrum Manipulation: By altering the frequency spectrum, attackers can create adversarial audio that evades detection, such as by using frequency compensation to counteract speaker-independent ASR systems.
- Acoustic Propagation Manipulation: In vehicle scenarios, attackers can manipulate audio inputs to ASR systems, potentially causing dangerous actions or accidents.
- Spectrum Reduction Attacks: By removing certain frequency components, attackers can create audio that is understandable to humans but not to machines, evading content moderation systems.
These vulnerabilities are exploited through various techniques, including modulated replay attacks, dolphin attacks, and voice masquerading, which can lead to unauthorized access, manipulation, and dissemination of harmful content.