The widespread adoption of Internet of Things (IoT) devices has made available a huge amount of data, thus enabling the development of applications that use machine learning (ML) to gather and dig up information, even from a user’s private life (such as sensors over the body). However, there are two major privacy preservation issues when using data to train and infer from ML: the small amount of computational resources in the devices, and the communication channels with the cloud infrastructure, which can have delays and may be intermittent. These two factors influence the architecture--should computation be done in the devices or in the cloud? And what about sensitive data carried by the devices? In fact, privacy preservation poses a dilemma: should user data be obfuscated before leaving the devices or in the cloud?
The authors provide a taxonomy of the existing privacy-preserving approaches, and for each technique they cover eventual attacks, shortcomings, and limitations. They also present ObfNet, a ML algorithm for preserving privacy at the initial stage.
This topic is of utmost importance nowadays. In fact, IoT devices (and even medical devices) share user data for secondary usage. Although regulations such as the General Data Protection Regulation (GDPR) provide a legal framework to protect users, proactive and “by design” privacy preservation is required.
Technical readers and IoT system architects will benefit from both the literature survey and the proposed algorithms.