DeepPatch: A Patching-Based Method for Repairing Deep Neural Networks

Deep neural networks (DNNs) have achieved competitive performance in many fields, but they also have defects to be repaired. For example, when a DNN encounters images with some corruption patterns (e.g. fog, Gaussian noise and glass blur) that it has never seen in the training dataset, its classification accuracy may decrease. A typical repair method is to add these corrupted images to the original training dataset and retrain the network. However, in many DNN repair scenarios, the original training dataset is not available, and retraining the network with only corrupted images may significantly decrease its accuracy on clean images. Therefore, without using the original training dataset, how to improve the network's accuracy on corrupted images while avoiding too much impact on clean images becomes an important issue. In this paper, we propose DeepPatch, a patching-based DNN repair method to tackle this issue. We retrain the original network with corrupted images to obtain a new network. In most cases, the new network performs better on corrupted images than the original network, and the original network performs better on clean images than the new network. To combine their advantages, we train an arbiter classifier. For an input image, the arbiter judges whether it is clean or corrupted. If it is clean, the original network is used for classification, otherwise the new network is used. We conduct large-scale experiments on 15 corruption patterns, four DNN models and three baselines. Experimental results show that our method achieves better performance than the baseline methods in most cases.