Applications of On-Die Neural Networks in Robust and Secure Analog/RF ICs and an Implementation in a Contemporary Technology




Journal Title

Journal ISSN

Volume Title



While manufactured ICs are subjected to extensive scrutiny in order to weed out defective or suspicious parts prior to their deployment, a variety of reasons such as silicon aging and adverse operational or environmental conditions might cause the performances of a previously healthy analog/RF IC to fail its design specifications. Similarly, field-activated triggers of hidden capabilities might cause a previously trusted analog/RF IC to exhibit malicious functionality. Additionally, prevention of unauthorized use of analog/RF ICs constitutes a major challenge in the field of hardware security. With analog/RF ICs now prevalent in most electronic systems, due to the rapid growth of wireless communications, sensor applications, and the Internet of Things (IoT), equipping them with post-deployment robustness, trustworthiness and performance locking mechanisms is very important. In this work, we demonstrate that on-die learning through an analog neural network can provide the aforementioned post-deployment capabilites. More specifically, the on-die neural network can be trained to (i) enhance robustness by calibrating the performances of an analog/RF IC, (ii) ensure trustworthiness by detecting the activation of potentially malicious circuitry, and/or (iii) prevent the unauthorized use of analog/RF ICs through performance locking. Based on the findings while implementing the three on-die learning tasks, we proceeded with the design, fabrication and fully characterization of an analog neural network in Globalfoundries’ 130nm RF CMOS process. Finally, all methods proposed in this dissertation have been verified with measurements from actual silicon chips.



Analog integrated circuits, Radio frequency integrated circuits, Robust control, Computer security


©2020 Georgios Volanis. All rights reserved.