Advances in SCA and RF-DNA Fingerprinting Through Enhanced Linear Regression Attacks and Application of Random Forest Classifiers

Radio Frequency (RF) emissions from electronic devices expose security vulnerabilities that can be used by an attacker to extract otherwise unobtainable information. Two realms of study were investigated here, including the exploitation of 1) unintentional RF emissions in the field of Side Channel Analysis (SCA), and 2) intentional RF emissions from physical devices in the field of RF-Distinct Native Attribute (RF-DNA) fingerprinting. Statistical analysis on the linear model fit to measured SCA data in Linear Regression Attacks (LRA) improved performance, achieving 98% success rate for AES key-byte identification from unintentional emissions. However, the presence of non-Gaussian noise required the use of a non-parametric classifier to further improve key guessing attacks. RndF based profiling attacks were successful in very high dimensional data sets, correctly guessing all 16 bytes of the AES key with a 50,000 variable dataset. With variable reduction, Random Forest still outperformed Template Attack for this data set, requiring fewer traces and achieving higher success rates with lower misclassification rate. Finally, the use of a RndF classifier is examined for intentional RF emissions from ZigBee devices to enhance security using RF-DNA fingerprinting. RndF outperformed parametric MDA/ML and non-parametric GRLVQI classifiers, providing up to GS =18.0 dB improvement (reduction in required SNR). Network penetration, measured using rogue ZigBee devices, show that the RndF method improved rogue rejection in noisier environments - gains of up to GS =18.0 dB are realized over previous methods