Analog-Domain Security Primitives for Internet-of-Things

In recent years, internet of things (IoT) has made tremendous progress and continues to evolve to create an ever more connected world by coalescing technology and human interaction. However, since IoT edge devices are often operated in unattended, hostile environments, they are significantly vulnerable to malicious attacks such as counterfeit IC usage, side-channel, and perception attacks. Moreover, severe resource constraints (limited energy, form factor, storage, bandwidth, computational resources) of edge-devices makes detection of such malicious attacks challenging. This thesis defense presents resource efficient CMOS analog/mixed-signal solutions to secure IoT edge devices. At first, to detect use of counterfeit ICs, an intrinsic and database-free IC authentication approach using an analog-to-digital converter (ADC) is presented. In this approach, since ADC is used for data-conversion as well as authentication, energy/area efficiency is improved by avoiding dedicated authentication modules. Next, this work presents co-design of low-power/area CMOS mixed-signal frameworks with statistical modeling methods for on-the-edge, real-time, perception attack and side-channel attack detection. Adopted statistical modeling approaches include non-parametric kernel density estimation (KDE) and parametric harmonic-mean of Gaussian-like Mixture Modeling (HMGM). Unlike neural network-based approaches, statistical modeling-based approaches are computationally light-weight and can be easily updated, hence, enables low latency, energy efficient and highly accurate attack detection. Subsequently, on-silicon implementation of kernel density approach using TSMC 65nm technology is discussed for fast statistical feature extraction from time-domain signals. Proposed feature-extraction approach is co-designed with deep-learning model for anomaly detection in time-series signals.