Taming Reachability Analysis of DNN-Controlled Systems via Abstraction-Based Training

Explore a novel approach to reachability analysis of deep neural network (DNN)-controlled systems presented at VMCAI'24. Learn about an abstraction-based training method that addresses the challenges of verifying complex DNNs and their hosting systems. Discover how inserting an additional abstraction layer during training allows for a blackbox reachability analysis approach that is sound, tight, efficient, and agnostic to DNN type and size. Examine the experimental results demonstrating comparable DNN performance and significant improvements in tightness and efficiency over existing white-box approaches. Gain insights into overcoming the limitations of current methods that rely on overapproximating DNNs using simpler polynomial models.