On the Challenge of Hardware Errors, Adversarial Attacks and Privacy Leakage for Embedded Machine Learning

Machine Learning deployment in Embedded Systems and Edge devices offer interesting advantages compared with the Cloud-based approaches, especially from a power consumption and environmental impact perspective. However, two principal problems need to be addressed towards trustworthy Embedded ML; first, Robustness to errors: several sources of faults can jeopardize ML systems integrity; be it hardware failures, as well as malicious fault injection. Second, Security and Privacy: this includes adversarial attacks and information leakage. In this chapter, we investigate these issues with an exploratory study on inherent fault tolerance of deep neural networks, as well as an overview on Embedded Systems-friendly defenses against adversarial attacks. Moreover, we provide an overview on privacy issues and discuss open problems we think the community needs to investigate.