Exploiting peer-to-peer communications for query privacy preservation in voice assistant systems

Voice assistant system (VAS) is a popular technology for users to interact with the Internet and the Internet-of-Things devices. In the VAS, voice queries are linked to users' accounts, resulting in long-term and continuous profiling at the service provider. In this paper, we propose a VAS anonymizer aiming to mix the queries of the VAS users to increase the source anonymity. The VAS anonymizer is equipped with a pattern-matching scheme, which allows VAS devices to find effective peer relays without disclosing their query patterns. Furthermore, the VAS anonymizer is equipped with anonymity evaluation modules for evaluating real-time single query, thus reducing the risk of pattern violation at the relays. Both the requester and the relay will evaluate the real-time query based on the resulting anonymity. Only if the anonymity evaluations at both requester and relay are positive, the query will be sent to the service provider via the relay. The VAS anonymizers at VAS devices coordinate the query uploading such that the sources of the queries are anonymized, and the service provider is unable to link the voice queries to individual users. In the experiments using our customized VAS devices and the Amazon Cloud servers, the computation and communication overhead of the matching scheme is shown to be efficient, and the anonymity evaluation modules are shown to be effective in protecting the privacy of the requesters and the relays.