Acoustic CAPTCHAs
HGI » UbiCrypt » Projects

Acoustic CAPT­CHAs

1st ad­vi­sor: Ko­los­sa, 2nd ad­vi­sor: Holz, 3rd ad­vi­sor: Schwenk

Gra­phi­cal CAPT­CHAs (Com­ple­te­ly Au­to­ma­ted Pu­blic Tu­ring tests to Tell Com­pu­ters and Hu­mans Apart) have been used with si­gni­fi­cant suc­cess for dis­tin­gu­is­hing human users from au­to­ma­ti­cal­ly per­for­med ac­ces­ses to net­work ser­vices, for ex­amp­le du­ring the re­gis­tra­ti­on pro­cess of on­line ser­vices. Acoustic CAPT­CHAs are re­qui­red for al­lowing ac­cess to vi­sual­ly im­pai­red users, but they have stub­born­ly re­mained much ea­sier to cir­cum­vent au­to­ma­ti­cal­ly. This the­sis will ex­plo­re two ave­nues of re­se­arch to im­pro­ve their se­cu­ri­ty:

In the first part, the cur­rent we­ak­nes­ses of sta­te-of-the-art au­to­ma­tic speech re­co­gni­ti­on (ASR) sys­tems will be ex­plo­red sys­te­ma­ti­cal­ly. The re­sult of this phase will be novel kinds of acoustic CAPT­CHAs that are more dif­fi­cult to crack than the cur­rent ones, while being equal­ly easy to un­der­stand for human lis­teners.

In the se­cond part of the pro­ject, the Ph.D. stu­dent will be wor­king on ques­ti­on sets for CAPT­CHAs which use spe­ci­fic human ca­pa­bi­li­ties - as­king ques­ti­ons re­qui­ring com­mon know­ledge or good na­tu­ral lan­gua­ge un­der­stan­ding skills which ma­chi­nes are still la­cking.

Over­all, the goal is the truly au­to­ma­tic de­sign of acoustic pseu­do Tu­ring tests, al­lowing the re­lia­ble dis­tinc­tion of human and ma­chi­ne and, along the way, gai­ning a bet­ter un­der­stan­ding of the spe­ci­fic dif­fe­ren­ces in per­cep­ti­on and co­gni­ti­on that dis­tin­giush com­pu­ters from hu­mans as of today.

Prof. Dorothea Kolossa

Prof. Dorothea Kolossa