Countering Modeling Attacks in PUF-based IoT Security Solutions

Hardware fingerprinting has emerged as a viable option for safeguarding IoT devices from cyberattacks. Such a fingerprint is used to not only authenticate the interconnected devices but also to derive cryptographic keys for ensuring data integrity and confidentiality. A Physically Unclonable Function (PUF) is deemed as an effective fingerprinting mechanism for resource-constrained IoT devices since it is simple to implement and imposes little overhead. A PUF design is realized based on the unintentional variations of microelectronics manufacturing processes. When queried with input bits (challenge), a PUF outputs a response that depends on such variations and this uniquely identifies the device. However, machine learning techniques constitute a threat where intercepted challenge-response pairs (CRPs) could be used to model the PUF and predict its output. This paper proposes an adversarial machine learning based methodology to counter such a threat. An effective label flipping approach is proposed where the attacker's model is poisoned by providing wrong CRPs. We employ an adaptive poisoning strategy that factors in potentially leaked information, i.e., the intercepted CRPs, and introduces randomness in the poisoning pattern to prevent exclusion of these wrong CRPs as outliers. The server and client use a lightweight procedure to coordinate and predict poisoned CRP exchanges. Specifically, we employ the same pseudo random number generator at communicating parties to ensure synchronization and consensus between them, and to vary the poisoning pattern over time. Our approach has been validated using datasets generated via a PUF implementation on an FPGA. The results have confirmed the effectiveness of our approach in defeating prominent PUF modeling attack techniques in the literature.