Repairing confusion and bias errors for DNN-based image classifiers

Recent works in DNN testing show that DNN based image classifiers are susceptible to confusion and bias errors. A DNN model, even robust trained model can be highly confused between certain pair of objects or highly bias towards some object than others. In this paper, we propose a differentiable distance metric, which is highly correlated with confusion errors. We propose a repairing approach by increasing the distance between two classes during retraining the model to reduce the confusion errors. We evaluate our approaches on both single-label and multi-label classification models and datasets. Our results show that our approach effectively reduce confusion errors with very slight accuracy reduce.