Performance Evaluation of Post-Quantum TLS 1.3 on Embedded Systems

Transport Layer Security (TLS) constitutes one of the most widely used protocols for securing Internet communication and has found broad acceptance also in the Internet of Things (IoT) domain. As we progress towards a security environment resistant against quantum computer attacks, TLS needs to be transformed in order to support post-quantum cryptography schemes. However, post-quantum TLS is still not standardized and its overall performance, especially in resource constrained, IoT capable, embedded devices is not well understood. In this paper, we evaluate the time, memory and energy requirements of a post-quantum variant of TLS version 1.3 (PQ TLS 1.3), by integrating the pqm4 library implementations of NIST round 3 post-quantum algorithms Kyber, Saber, Dilithium and Falcon into the popular wolfSSL TLS 1.3 library. In particular, our experiments focus on low end, resource constrained embedded devices manifested in the ARM Cortex-M4 embedded platform NUCLEO-F439ZI (with hardware cryptographic accelerator) and NUCLEO-F429ZI (without hardware cryptographic accelerator) boards. These two boards only provide 180MHz clock rate, 2 MB Flash Memory and 256 KB SRAM. To the authors’ knowledge this is the first thorough time delay, memory usage and energy consumption PQ TLS 1.3 evaluation using the NIST round 3 finalist algorithms for resource constrained embedded systems with and without cryptography hardware acceleration. The paper’s results show that the post-quantum signatures Dilithium and Falcon and post-quantum KEMs Kyber and Saber perform in general well in TLS 1.3 on embedded devices in terms of both TLS handshake time and energy consumption. There is no significant difference between the TLS handshake time of Kyber and Saber; However, the handshake time with Falcon is much lower than that with Dilithium. In addition, hardware cryptographic accelerator for symmetric-key primitives improves the performances of TLS handshake time by about 6% on the client side and even by 19% on the server side, on high security levels.