Hiding Sensitive Information in Desensitized Voice Sequences

Voice data is broadly acquired and utilized by consumer services. In order to process such data, most of the raw records are sent to web servers, possibly with dedicated acceleration hardware. However, in this way malicious service providers can identify the users because the raw voice sequences contain rich voiceprint information, which is adequate for deduction of a large amount of private information. In order to mitigate such problem, desensitization methods are employed as secure intermediaries between user and the cloud services. However, if these methods are provided by a third party as a black box, it may not proved to be safe enough. In this paper, we demonstrate and experiment the possibility of hiding information sufficient to extract original voice from in seemingly desensitized voices that may be used for various online services, utilizing StarGAN-based voice transformation and voice-optimized audio stenography technologies.