Developing robust sparse models fit for safety-critical and resource-constrained systems such as drones, autonomous robots, etc., has been an issue of longstanding interest. The inability of adversarial training mechanisms to provide a formal robustness guarantee kindles the requirement for verified local robustness mechanisms. This work aims to compute sparse verified locally robust networks which exhibit (benign) accuracy and verified local robustness comparable to their dense counterparts. Towards this objective, we examine several model sparsification approaches and present `SparseVLR'-- a framework to search verified locally robust sparse networks. We empirically investigated SparseVLR's efficacy and generalizability by evaluating various benchmark and application-specific datasets across several models. Above all, we provide an in-depth study and reasoning to unveil the causes for the ascendancy of SparseVLR.
View on arXiv