Approximate computing has emerged as an important area of research and practice, driven by the advent of artificial intelligence and new approaches to data processing.
While the field of approximate computing is very broad, many subfields of computing have developed robust resilience to errors, but some other areas, from quantum materials to healthcare, cannot always function with a significant loss of accuracy.
Security was considered one of the areas where accuracy is of paramount importance. But recently interest in approximate computing for security has entered the research arena. It is motivated in part by the requirements for security in systems relying on approximate computing and the importance of energy efficiency and optimized system performance.
A group at the Queens University Belfast (QUB) has examined challenges and opportunities of approximate computing for security in a paper published in 2020[1]. The paper identifies some areas relevant to the discussion of approximate computing in the security space based on threat analysis. In hardware, important threats include side channel attacks, reverse engineering, hardware trojans or logic obfuscation. Cryptography is undergoing a broad scale revision with the expected advent of quantum computers and the rise of analytics on encrypted data. In network security, more complex network topology, modularization, and very high bandwidth changed the requirements for network resilience with the introduction of 5G and the expected entry of 6G around 2030. In the modern computing environment, approximation can create additional security challenges and escalate existing attacks, as well as increasing the impact from attacks that were not previously considered practical. This threat landscape creates additional challenges for defenders of complex systems with multiple elements, especially when approximate computing is in play.
But the use of approximate computing models can also create new opportunities for security and defenders of systems. The QUB paper puts forward, in general terms, several areas in security or adjacent to security where approximation brings benefits. Some subfields of security are ahead in using approximate computing models. In biometrics, approximation is commonly used and was essential for creating commercial systems in, for example, border control or to enhance security of portable devices, like smart phones.
In recent literature, models taking advantage of approximation were proposed for encryption. For example, the same QUB group described an area/power efficient approximate modular multiplier (so called AxMM) for complete RLWE (ring learning with error) hardware based on approximation principles.[2] Several homomorphic encryption schemes using approximate arithmetic based on RLWE were also devised; for instance, the scheme described by Cheon et al[3]. Approximation-based systems for memory (BRAM and MRAM) were also developed and seem promising due to their energy efficiency, for example, as elucidated by Li et al.[4]
Will the use of approximate computing become more pervasive in the security space, beyond areas like biometrics where it is already common? There are many challenges on this path, despite considerable progress made already. It remains a promising area of research.
[1] Liu, Weiqiang, et al. "Security in approximate computing and approximate computing for security: Challenges and opportunities." Proceedings of the IEEE 108.12 (2020): 2214-2231.
[2] Y. Zhang, C. Wang, D. E. S. Kundi, A. Khalid, M. O’Neill, and W. Liu, “An efficient and parallel R-LWE cryptoprocessor,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 67, no. 5, pp. 886–890, 2020
[3] J. H. Cheon, A. Kim, M. Kim, and Y. Song, “Homomorphic encryption for arithmetic of approximate numbers,” in Proc. International Conference on the Theory and Application of Cryptology and Information Security, 2017, pp. 409–437
[4] B. Li, P. Gu, Y. Wang, and H. Yang, “Exploring the precision limitation for RRA M-based analog approximate computing,” IEEE Design & Test, vol. 33, no. 1, pp. 51–58, 2015.