Privacy-Preserving Speaker Verification as String Comparison

Abstract

In this chapter, we develop a method for speaker verification that requires minimal computation overhead needed to satisfy the privacy constraints. The central aspect of our approach is to reduce the speaker verification task to string comparison. Instead of using the UBM-GMM approach, we convert the utterances into supervector features (Campbell et al., 2006) that are invariant with the length of the utterance. By applying the locality sensitive hashing (LSH) transformation (Gionis et al., 1999) to the supervectors, we reduce the problem of nearest-neighbor classification into string comparison. It is very efficient to perform string comparison with privacy, similar to a conventional password system. By applying a cryptographic hash function, e.g., SHA-256 (SHA 2008), we convert the LSH transformation to an obfuscated string which the server cannot use to gain information about the supervectors, but is still able to compare if two strings are identical. This one-way transformation preserves the privacy of the speech utterances submitted by the user, and can be executed significantly faster than applying homomorphic encryption.