DivNet
DivNet
ACM MobiSys 2026

DivNet

Ying Meng and Qiang Zeng

George Mason University

What are the threats?

Audio adversarial examples are maliciously crafted audio clips that sound normal to human listeners yet cause automatic speech recognition (ASR) systems to produce attacker-intended outputs. Such attacks pose serious real-world threats, including hijacking smarthome devices, triggering unauthorized transactions, falsifying meeting transcripts, and manipulating voice-controlled vehicles.

An audio clip streamed from the internet is embedded with a hidden command ("open the front door"). It sounds like ordinary music to human listeners, yet the smart-home device interprets the concealed command as legitimate and unlocks the door.
An audio clip is embedded with a hidden command ("transfer one thousand dollars") and played out loud near a person who is talking to their smart banking assistant. The clip sounds like ordinary music to human listeners, yet the assistant interprets the concealed command as a legitimate instruction and carries out an unauthorized money transfer.

Samples

The same audio is fed to each deployment: the original (the attack target) and two of its variants. On benign audio all three recognize correctly and accurately. On adversarial audio, only the original model is fooled into the attacker's target phrase — each variant hears something else, and the two variants disagree with each other.

Benign audio

Normal speech. Every model transcribes it the same way.

Audio Original model Variant 1 Variant 2
HE COULD WAIT NO LONGER HE COULD WAIT NO LONGER HE COULD WAIT NO LONGER
YOUR POWER IS SUFFICIENT I SAID YOUR POWER IS SUFFICIENT I SAID YOUR POWER IS SUFFICIENT I SAID

CW adversarial examples

Gradient-based attack. The original model is hijacked; the variants are not.

Audio Original model Variant 1 Variant 2
A MAN IN THE WELL Y MY IN A I MA INSA WELL
YES MANY TIMES unnot recognized YES A ANE KINES
I KNOW NOW WHAT BRINGS ME TO THE FIRE YOUR MOTHER THE QUEEN MA CAN E I YOUR MOTHER THE QUEEN Y STANDI BY
WHERE IS MY HUSBAND THERE IS ME AEN WHERE IS MEALS TEN
I DID NOT KNOW WHAT HE MEANT I TEN E MONY MEN A DREADEUM WHAT HE MENT

PHVC adversarial examples

Signal-processing hidden voice commands. Only the original obeys the command.

Audio Original model Variant 1 Variant 2
TURN ON ALL THE LIGHTS not recognized CARE O THE LICS
AH WHAT SHALL WE GO FOR A HOLE not recognized I WATHA OVER WHOLE

KENKU adversarial examples

A state-of-the-art transferable attack. The variants still hold.

Audio Original model Variant 1 Variant 2
WHAT IS THE WHEATHER TODAY not recognized DAY
UNLOCK THE DOORS not recognized ORS

ALIF adversarial examples

Linguistic-feature attack with high reported transferability. Variants diverge.

Audio Original model Variant 1 Variant 2
GOOD MORNING not recognized not recognized
CLEAR NOTIFICATION ER NEFICATION HOIL NOFICATION