Trained with a sufficiently large training and testing dataset, Deep Neural Networks (DNNs) are expected to generalize. However, inputs may deviate from the training dataset distribution in real deployments. This is a fundamental issue with using a finite dataset, which may lead deployed DNNs to mis-predict in production.

Inspired by input-debugging techniques for traditional software systems, we propose a runtime approach to identify and fix failure-inducing inputs in deep learning systems. Specifically, our approach targets DNN mis-predictions caused by unexpected (deviating and out-of-distribution) runtime inputs. Our approach has two steps. First, it recognizes and distinguishes deviating (``unseen'' semantically-preserving) and out-of-distribution inputs from in-distribution inputs. Second, our approach fixes the failure-inducing inputs by transforming them into inputs from the training set that have similar semantics. We call this process \emph{input reflection} and formulate it as a search problem over the embedding space on the training set.

We implemented a tool called InputReflector based on the above two-step approach and evaluated it with experiments on three DNN models trained on CIFAR-10, MNIST, and FMNIST image datasets. The results show that InputReflector can effectively distinguish deviating inputs that retain semantics of the distribution (e.g., zoomed images) and out-of-distribution inputs from in-distribution inputs. InputReflector repairs deviating inputs and achieves 30.78% accuracy improvement over original models. We also illustrate how InputReflector can be used to evaluate tests generated by deep learning testing tools.