What is it about?
The BackDoor exploit is a way to make microphones record sound which is inaudible to humans. To a human it can be very difficult to tell if a recording was actually made audibly or if it was inserted inaudibly by the BackDoor exploit. By using deep learning, it is possible to analyze subtle aspects of the recording which can allow us to determine whether the recording originated from BackDoor or not. Our preliminary results indicate that this is possible, and provide some associated metrics as well as areas for further improvement.
Featured Image
Photo by Sebastian Scholz (Nuki) on Unsplash
Why is it important?
Certain Voice-Enabled Devices (VEDs) may be manipulated through the BackDoor system into performing actions which its user is unaware of and which are inaudible to humans, making it difficult to tell how such instructions are being given. By using deep learning, it may be possible to equip such VEDs with a built-in system to defend against these attacks by detecting the use of BackDoor.
Read the Original
This page is a summary of: Detecting acoustic backdoor transmission of inaudible messages using deep learning, July 2020, ACM (Association for Computing Machinery),
DOI: 10.1145/3395352.3402629.
You can read the full text:
Contributors
The following have contributed to this page
