AI-Based Side Channel Attack for EW Cyber Offense and Defense

Unlike the previous and existing works, the Phase II effort will focus on anti-tampering, reverse engineering, obfuscation, and SCA hardening through plug-and-play. The core component is the deep neural network-based model for EM emission traces.  We plan to investigate SCA for EM and RF with deep neural networks (DNNs) at the architecture/circuit level, test SCA resiliency on existing embedded systems, and establish digital twins and test protocols for SCA hardenings.  We plan to integrate side-channel attack testing, fault Injections, cryptanalysis, electromagnetic (EM) spectrum analysis, wireless & wired Communication, and digital communication protocols in a set of generative pre-training models in the DNNs to guide automatic and secure embedded system designs. The technologies offered by Black Fur are unique: we develop DNNs that extend the cryptanalysis capabilities with real-time signals, power traces, EM spectrums, and other features that pertain to system performance. As is well known, SCAs are divided into profiled and non-profiled categories. In a non-profiled assault, an adversary constructs a side-channel leakage model using knowledge of the target system. Once an adversary obtains the model, they can derive the secret key by correlating the side channel from the target and the leakage model. The effectiveness of a non-profiled EM leakage model depends on device architecture, software implementation, and manufacturing technology characteristics. Profiling side-channel attacks, such as template attacks, rely on total leakage to characterize the target device. An adversary can profile the target device and create a leakage template for every conceivable key. The robustness of profiled SCAs to system disturbance and countermeasures is enhanced.  Our recent research has investigated the inter-device portability of DNNs for SCAs across different physical manifestations of the various device architectures that potentially can be applied for EM and RF applications. They demonstrated that manufacturing variations and environmental conditions can significantly impact other devices' efficacy. They enhanced inter-device performance by combining device data before training. Notably, this portability addresses the physical differences between devices with the inter-architecture portability between devices with various processor architecture configurations.