Privacy Preserving Techniques for Speech Processing

Speech is perhaps the most private form of personal communication but current speech processing techniques are not designed to preserve the privacy of the speaker and require complete access to the speech recording. We propose to develop techniques for speech processing which do preserve privacy. While our proposed methods can be applied to a variety of speech processing problems and also generally to problems in other domains, we focus on the problems of keyword spotting, speaker identification, speaker verification, and speech recognition. Each of these applications involve the two separate but overlapping problems of private classifier evaluation and private classifier release. Towards the former, we study designing privacy preserving protocols using primitives such as homomorphic encryption, oblivious transfer, and blind and permute in the framework of secure multiparty computation (SMC). Towards the latter, we study the differential privacy model, and techniques such as exponential method and sensitivity method for releasing differentially private classifiers trained on private data. We summarize our preliminary work on the subject including techniques such as SMC protocols for eigenvector computation, isolated keyword recognition and differentially private release mechanisms for large margin Gaussian mixture models, and classifiers trained from data belonging to multiple parties. Finally, we discuss the proposed directions of research including techniques such as training differentially private hidden Markov models, and a multiparty framework for differentially private classification by data perturbation, and protocols for applications such as privacy preserving music matching and keyword spotting, speaker verification and identification.