May_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 2 34 BETTER PERFORMANCE, HIGHER RELIABILITY, MORE SECURITY: RESEARCH HIGHLIGHTS FROM THE CENTER FOR ADVANCED ELECTRONICS THROUGH MACHINE LEARNING Aydin Aysu, Xu Chen, W. Rhett Davis, Sung Kyu Lim, Paul Franzon, Madhavan Swaminathan, and Elyse Rosenbaum elyse@illinois.edu UNIVERSITY HIGHLIGHT T he Center for Advanced Electronics through Machine Learning, or CAEML, is an Industry- University Cooperative Research Center (IUCRC) supportedby theNational Science Foundation. The IUCRC programsupports pre-competitive research that is of high interest to industry. CAEML’s research mission is to utilize machine learning for the design of optimized microelec- tronic circuits and systems and to increase the efficiency of electronic design automation. The center’s work is carried out by researchers at the University of Illinois Urbana-Champaign, Georgia Tech, and North Carolina State University. As part of NSF’s IUCRC program, CAEML is additionally tasked with building the workforce by assigning university students to conduct industry-relevant research and helping those students to become skilled at collaboration and communication. Over the past five years, CAEML has carried out re- search on applications ranging from security to physical design, and from analog design to signal integrity. A few of the center’s research projects are described here. MACHINE LEARNING ENHANCED SIDE-CHANNEL ATTACKS Although cryptographic systems may be theoretically secure against mathematical attacks, their practical real- izations in hardware and software can leak information. One such popular attack vector on cryptographic imple- mentations is side-channel analysis, which aims to extract secret cryptographic keys through correlating processing by-products such as electromagnetic radiation, power draw, or execution time to the secret-data dependent computations. Two decades of research on side-channel analysis have shown that such attacks are indeedpractical on a range of devices fromedge/IoTnodes to cloud servers and that countermeasures must be employed tomitigate their vulnerability. A team led by Prof. Aydin Aysu at NC State demonstrat- ed that the classical techniques developed in the last two decades can be drastically improved by utilizing recent advances in machine learning. The team demonstrated that a convolutional neural network, when trained with the power consumption measurements of a device, can automatically identify the secret data being processed on a new device with unknown keys. [1] This means that an adversary can “hack” into a system with the assistance of machine learning. The process, however, is nontrivial and requires adapting data augmentation techniques that translate one-dimensional power consumption data into two-dimensional images on which the neural network excels. The team’s findings also revealed that neural networks can learn the effects of obfuscation-based defenses andnegate them, effectively evadingwell-known defenses against side-channel attacks. By contrast, clas- sical techniques with limited statistical analysis failed to detect the vulnerability and would mistakenly conclude that the implementation is secure unless the adversary THE FDTD METHOD IS A WELL- ESTABLISHED APPROACH TO SIMULATING THE TIME EVOLUTION OF ELECTROMAGNETIC FIELDS.