CommanderGabble: A Universal Attack Against ASR Systems Leveraging Fast Speech
Introduction
In this project, we discover that fast speech can be utilized to camouflage adversarial voice commands against Automatic Speech Recognition (ASR) systems. We refer a novel technique refer to as CommanderGabble.
Adversarial Audio Examples
We provide example advarsary audios with the corresponding attack results or commands.
We test commands with three differnet lengths:
Short: One word
Medium: 2-3 words
Long: More than 3 words
Over-the-wire attack example
Audio 1 (short):
"Louder"
Audio 2 (short):
"Enable"
Audio 3 (medium):
"Play News"
Audio 4 (medium):
"Turn Up Volume"
Audio 5 (long):
"Where is My Stuff"
Audio 6 (long):
"Delete This Song from Playlist"
Over-the-air attack example
Household environment
Audio 7 (short):
"Pause"
Audio 8 (medium):
"Call My Phone"
Audio 9 (long):
"Turn Off Light in Living Room"
Teleconference environment
Audio 10 (short):
"Bluetooth"
Audio 11 (medium):
"Recent Messages"
Audio 12 (long):
"Set Alarm at 3 am"
Vehicle environment
Audio 13 (short):
"News"
Audio 14 (medium):
"Cancel Route"
Audio 15 (long):
"How Long Will It Take to Drive to Library"
Inventory List
Below are the artifacts that generate the main results in our paper.
1. Pass rate tests for CVC-contained and CVC-free words [link]
Pass rate test result for CVC-contained and CVC-free words (Figure 3)