Source distinguishability under corrupted training | IEEE Conference Publication | IEEE Xplore