Search

Smart speakers and voice-based virtual assistants are core components for the success of the IoT paradigm. Unfortunately, they are vulnerable to various privacy threats exploiting machine learning to analyze the generated encrypted traffic. To cope with that, deep adversarial learning approaches can be used to build black-box countermeasures altering the network traffic (e.g., via packet padding) and its statistical information. This letter showcases the inadequacy of such countermeasures against machine learning attacks with a dedicated experimental campaign on a real network dataset. Results indicate the need for a major re-engineering to guarantee the suitable protection of commercially available smart speakers.

Smart speakers and voice-based virtual assistants are used to retrieve information, interact with other devices, and command a variety of Internet of Things (IoT) nodes. To this aim, smart speakers and voice-based assistants typically take advantage of cloud architectures: vocal commands of the user are sampled, sent through the Internet to be processed and transmitted back for local execution, e.g., to activate an IoT device. Unfortunately, even if privacy and security are enforced through state- of-the-art encryption mechanisms, the features of the encrypted traffic, such as the throughput, the size of protocol data units or the IP addresses can leak critical information about the habits of the users. In this perspective, in this paper we showcase this kind of risks by exploiting machine learn- ing techniques to develop black-box models to classify traffic and implement privacy leaking attacks automatically. We prove that such traffic analysis allows to detect the presence of a person in a house equipped with a Google Home device, even if the same person does not interact with the smart device. We also present a set of experimental results collected in a realistic scenario, and propose possible countermeasures.