TinyML

Chapter 20. Privacy, Security, and Deployment

After working through the previous chapters in this book, you should hopefully be able to build an embedded application that relies on machine learning. You’ll still need to navigate a lot of challenges, though, to turn your project into a product that can be successfully deployed into the world. Two key challenges are protecting the privacy and the security of your users. This chapter covers some of the approaches we’ve found useful for overcoming those challenges.

Privacy

Machine learning on-device relies on sensor inputs. Some of these sensors, like microphones and cameras, raise obvious privacy concerns, but even others, like accelerometers, can be abused; for example, to identify individuals from their gait when wearing your product. We all have a responsibility as engineers to safeguard our users from damage that our products can cause, so it’s vital to think about privacy at all stages of the design. There are also legal implications to handling sensitive user data that are beyond the scope of our coverage but about which you should consult your lawyers. If you’re part of a large organization, you might have privacy specialists and processes that can help you with specialist knowledge. Even if you don’t have access to those resources, you should spend some time running your own privacy review at the outset of the project, and periodically revisit it until you launch. There isn’t widespread agreement on what a “privacy review” ...