Dictionary Attacks on Speaker Verification

In this paper, we propose dictionary attacks against speaker verification-a novel attack vector that aims to match a large fraction of population by chance. We introduce generic formulation the can be used with various speech representations and threat models. The attacker uses adversarial optimization maximize raw similarity embeddings between seed sample proxy population. resulting master voice successfully matches non-trivial people in an unknown Adversarial waveforms obtained our approach on average 69% females 38% males enrolled target system at strict decision threshold calibrated yield false alarm rate 1%. By using black-box cloning system, obtain voices are effective most challenging conditions transferable encoders. also show that, combined multiple attempts, opens even more serious issues security these systems.