引導式測驗 · 硬 等級

對語音辨識測驗的對抗性攻擊

Test adversarial versus ordinary ASR errors, digital and physical threat models, attack-transfer limits, defense tradeoffs, and confirmation.

問題 1 的 8

What distinguishes an adversarial audio example from accidental background noise?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. What distinguishes an adversarial audio example from accidental background noise?
    • An accidental background-noise burst
    • A routine codec artifact not chosen to alter recognition
    • It is intentionally crafted to induce a particular model error
    • An ordinary speaker mispronunciation
  2. What was the scope of the cited 2018 targeted-attack demonstration?
    • A particular recognizer under a white-box digital test
    • Reliable control of every current voice assistant
    • Guaranteed success through all rooms and speakers
    • A proof that people could not hear any change
  3. Why can digital-file attack success fail through a loudspeaker and microphone?
    • The file becomes a different language by definition
    • All microphones delete sound
    • Transcripts cannot be produced from playback
    • The physical acoustic channel changes the waveform
  4. Which threat-model detail matters before a voice system is evaluated?
    • Only the speaker’s accent distribution
    • What input and model access an attacker can obtain
    • Only the model’s architecture name
    • Only the average word count of recordings
  5. Why measure normal-user errors after adding an audio defense?
    • Defense success automatically improves recognition
    • Every normal utterance is an attack
    • A defense can reject or damage legitimate speech
    • Transcription accuracy is irrelevant to users
  6. How should a high-impact voice command be handled?
    • Require appropriate confirmation or authorization beyond a transcript
    • Execute whenever words appear with high confidence
    • Treat a transcript as proof of identity
    • Disable all audit logs
  7. Which claim is unjustified from one historical model’s vulnerability?
    • That adversarial ASR research exists
    • That threat models differ
    • That physical transfer needs testing
    • That a different current product is compromised without its own evidence
  8. What should be compared in an authorized robustness study?
    • Only attack success on one selected clip
    • Attack success and ordinary accuracy across relevant devices and rooms
    • Only clean speech accuracy with no attack trials
    • Only how similar the audio sounds to the attacker