MADEIRA: Multi-Agent Debloating Environment to Increase Robustness in Applications

SIFT and the University of Southern California's Information Sciences Institute (USC/ISI) propose to develop the Multi-Agent Debloating Environment to Increase Robustness in Applications (MADEIRA). MADEIRA will apply AI/ML techniques to learn how to apply diverse debloating tools, automatically reducing the attack surfaces of full systems that may include firmware, OS, container, and application-level targets. We will build on SIFT's proven Cyber Reasoning System (CRS), a fully autonomous, highly parallel plug-in architecture that already applies a wide variety of software analysis and rewriting tools to complex systems. We will also incorporate methods from SIFT's related OPENSHELL system, which uses massively parallel Symbolic Reinforcement Learning methods to learn how to apply different tools. USC/ISI brings a wealth of ML techniques for extracting knowledge from software, in both binary and source code forms. Designed to maximize the automation of the debloating process, MADEIRA will learn from successful debloating efforts on one target system and apply them to new targets. Ā ĀBy automatically assessing and validating the debloating results, MADEIRA will ensure correct application of the debloating tools, and will explore diverse settings to maximize security and integrity metrics.