Uncertainty for Anti-Spoofing

Bayesian neural networks that say how confident they are when deciding whether a speaker-verification attempt is spoofed.

Speech Communication 137, 44–51 · 2022 C. Suslu, Eray Eren, C. Demiroglu
Bayesianuncertainty estimation
ASVspoofing countermeasure
Second authormy contribution

A spoofing countermeasure that outputs only a hard accept/reject hides something important: how sure it actually is. That matters most exactly where these systems are weakest, on attack types they were never trained against.

This work applies a Bayesian treatment to spoofing detection for automatic speaker verification, so the model produces a calibrated sense of its own uncertainty alongside the decision. That estimate can then gate the outcome, flagging low-confidence cases instead of silently guessing.

Work led by Cagil Suslu at Ozyegin University; I contributed as second author.

Citation

@article{suslu2022uncertainty,
  title   = {Uncertainty assessment for detection of spoofing attacks to speaker verification systems using a Bayesian approach},
  author  = {Suslu, Cagil and Eren, Eray and Demiroglu, Cenk},
  journal = {Speech Communication},
  volume  = {137},
  pages   = {44--51},
  year    = {2022},
  doi     = {10.1016/j.specom.2021.12.003}
}