On optimization of hardware-assisted security

Abstract

Physically Unclonable Functions (PUFs) have emerged as lightweight hardware security primitives to implement authentication and key generation features on electronic devices. An ideal Strong PUF cannot be cloned and maps an n-bit input to a unique m-bit output. However, real Strong PUF implementations su?er from security issues. This thesis proposes various novel Strong PUF designs, based on Weightless Neural Network (WNN) architecture, which are resistant against model building attacks through machine learning algorithms. Then, the proposed WNN PUFs are combined with a reliable entropy source to extend the reliability property to the final Strong PUF. High volume manufacturing of PUFs requires online testing techniques to ensure the desired uniqueness property among the manufactured PUFs. An online testing PUF solution based on Multi-Index Hashing (MIH) is optimized by similarity search strategies to reduce the memory resources. Dynamic Information Flow Tracking (DIFT) has been successfully utilized to detect illegal access to sensitive information at runtime. Nonetheless, recent evasion attacks explore implicit flows based on control dependencies that are not detectable by most of DIFT implementations, which only track data dependency propagation. In this thesis, a portable nested implicit flow tracking is proposed to enable explicit flow based DIFT mechanisms track implicit flows, including deeply-nested branch scenarios. In addition, a new propagation rule is defined to mitigate the incorrect propagation of data under control dependencies. Finally, new WNN models based on probabilistic data structures are proposed and analyzed in order to reduce the memory requirements. The new models are robust and are suitable as components for hardware-assisted security solutions.