VeriSparse: Training Verified Locally Robust Sparse Neural Networks from Scratch

Several safety-critical applications such as self-navigation, health care, and industrial control systems use embedded systems as their core. Recent advancements in Neural Networks (NNs) in approximating complex functions make them well-suited for these domains. However, the compute-intensive nature of NNs limits their deployment and training in embedded systems with limited computation and storage capacities. Moreover, the adversarial vulnerability of NNs challenges their use in safety-critical scenarios. Hence, developing sparse models having robustness guarantees while leveraging fewer resources during training is critical in expanding NNs' use in safety-critical and resource-constrained embedding system settings. This paper presents 'VeriSparse'-- a framework to search verified locally robust sparse networks starting from a random sparse initialization (i.e., scratch). VeriSparse obtains sparse NNs exhibiting similar or higher verified local robustness, requiring one-third of the training time compared to the state-of-the-art approaches. Furthermore, VeriSparse performs both structured and unstructured sparsification, enabling storage, computing-resource, and computation time reduction during inference generation. Thus, it facilitates the resource-constraint embedding platforms to leverage verified robust NN models, expanding their scope to safety-critical, real-time, and edge applications. We exhaustively investigated VeriSparse's efficacy and generalizability by evaluating various benchmark and application-specific datasets across several model architectures.