Sandia National Labs FY20 LDRD Annual Report


Multimodal deep learning models improve automatic identification of security flaws in software. Sandia and its customers will operationalize machine learning-based methods that simultaneously leverage both source code and machine binary representations of software programs to identify security flaws, thanks to a recent LDRD project. Typically, software developed and used in national security missions is audited and tested before deployment by analyzing source code or reverse engineering a program’s binary representation. These newly developed machine learning models, based on multimodal deep learning approaches, were designed to model and analyze both representations simultaneously. Results of applying these models to the Juliet Test Suite, a benchmark used by researchers to demonstrate the efficacy of new methods, and a version of the Linux operating system kernel containing synthetically injected flaws, illustrate a marked performance improvement over existing state-of-the-art machine learning methods. (PI: Daniel Dunlavy)

Multimodal deep learning models, based on (a) correlation networks, (b) joint autoencoders, and (c) bidirectional deep neural networks, can accurately model security flaws in software by leveraging both source code and machine binary representations of programs.

Capability development for prediction and optimization of engineered anisotropic thermal barriers. Advanced barriers are needed to protect components from mechanical, thermal, electrical, chemical, electromagnetic, and radiation insults. Multi-layered or composite barriers may be used to manage different environments. The team developed new experimental capabilities to measure the anisotropic thermal properties of advanced thermal barriers. This capability will be the foundation for any further development of thermal model and material designs for advanced thermal barriers. (PI: Karla Reyes) Next-gen high-power electronic systems to be enabled by first AlGaN-based vertical power- switching transistors. The first aluminum gallium nitride- (AlGaN) based vertical power-switching transistor was the product of this Sandia research. The team developed a mechanistic understanding of how crystalline defects at a regrown p/n junction interface influence reverse electrical leakage in AlGaN-based diodes. This understanding enabled selective-area-regrown p/n junctions with low leakage current and will be used as the core element necessary to demonstrate the first AlGaN- based vertical power-switching transistor, an enabling capability for next-generation high-power electronic systems. (PI: Andrew Allerman)



Made with FlippingBook - professional solution for displaying marketing and sales documents online