Embedded devices increasingly rely on AI models to perform disparate tasks. Companies need to spend a lot of effort, time, and money to build AI models, such as deep neural networks, due to a long process that includes several onerous design and development phases. AI models are however susceptible to theft, tampering, and unauthorized access due to the vulnerabilities of the hardware and software environments. Concerning theft, when no proper countermeasurs are in place, attackers could steal intellectual property that required enormous investments to be generated.

Protecting the AI models on such devices while maintaining a seamless inference workflow is hence a major challenge.