Attacking and Defending Behind A Psychoacoustics-Based Captcha

This paper proposes a novel audio CAPTCHA system that requires a user to respond immediately after hearing a short and easy-to-remember cue in its mixture with background music. Potential attacking paths based on cross correlation (CC) and sound event detection (SED) are implemented to test the security of the system. Then, two defending measures based on phase-modification of the cue and audio watermarking are established against CC and SED-based attacks, respectively. Human subjects were recruited to test the system and the results indicated that the subjects can pass the proposed audio CAPTCHA >90% of the times. In contrast, the proposed defending measures suppress the passing rate of both attacks to 8.2% and 0.4%, respectively.