PART 6 · MODULE 6Attacking speech AI
Audio Adversarial Attacks
Adversarial attacks on audio models
Loading module...
Adversarial attacks on audio models
Just as images can be subtly modified to fool vision models, audio can be manipulated to fool speech recognition systems. The attacker adds noise that is inaudible to humans but completely changes what the AI "hears."
Imagine someone playing a song in a coffee shop. You hear music. But your voice assistant hears "send all my contacts to this email address."
Audio Waveform with Adversarial Perturbation
Speech recognition models convert audio into visual spectrograms (frequency over time). Attackers manipulate these frequency patterns to inject hidden commands that map to specific words in the model's vocabulary.
Human hearing has blind spots. Loud sounds mask quiet sounds at nearby frequencies. Attackers embed commands in these perceptual gaps where humans literally cannot hear them but the AI processes them perfectly.
These are not just digital tricks. Adversarial audio can be played through physical speakers and still fool nearby voice assistants. The attack survives the real-world journey from speaker to microphone.
An attacker plays adversarial audio near a smart speaker. The owner hears nothing unusual, but the assistant receives a command to make a purchase, unlock a door, or read out private messages.
Adversarial commands hidden in YouTube videos, podcasts, or music tracks. Anyone listening with a voice-activated device nearby could have their assistant hijacked without knowing.
Defenses include input verification (checking that audio matches expected human speech patterns), multi-model verification (using multiple different models that are hard to fool simultaneously), and user confirmation for sensitive actions.
Audio AI can be fooled by sounds invisible to human ears. As voice assistants control more of our lives — smart homes, cars, banking — inaudible adversarial commands become a serious privacy and security threat.
For hands-on practice, check out the AI Security course on GitHub.