Toward Continuous Verification of DNNs

Deep neural network (DNN) verification is used to assess whether a DNN meets a desired trustworthy property (e.g., robustness, fairness) on an infinite input set. While there have been significant advancements in scalable verifiers for individual DNNs, they suffer from inefficiency when developers update a deployed DNN (e.g. through quantization, pruning, or fine-tuning) to enhance its inference speed or accuracy. The verification is inefficient because the developers need to rerun the computationally expensive verifier from scratch on the updated DNN. To address this issue, we propose an incremental DNN verification framework, leveraging novel theory, data structures, and algorithms. Our previous works anonymous (2023) and anonymous (2022) focused on incremental verification, and in this paper, we provide a summary of these advancements and explore the potential of incremental verification in enabling real-world DNN verification systems.