The Triangular Trade-off between Robustness, Accuracy and Fairness in Deep Neural Networks: A Survey

Abstract

With the rapid development of deep learning, AI systems are being used more in complex and important domains and necessitates the simultaneous fulfillment of multiple constraints: accurate, robust, and fair. Accuracy measures how well a DNN can generalize to new data. Robustness demonstrates how well the network can withstand minor perturbations without changing the results. Fairness focuses on treating different groups equally. This survey provides an overview of the triangular trade-off among robustness, accuracy, and fairness in neural networks. This trade-off makes it difficult for AI systems to achieve true intelligence and is connected to generalization, robustness, and fairness in deep learning. The survey explores these trade-offs and their relationships to adversarial examples, adversarial training, and fair machine learning. The trade-offs between accuracy and robustness, accuracy and fairness, and robustness and fairness have been studied to different extents. However, there is a lack of taxonomy and analysis of these trade-offs. The accuracy-robustness trade-off is inherent in Gaussian models, but it varies when classes are not closely distributed. The accuracy-fairness and robustness-fairness trade-offs have been assessed empirically, but their theoretical nature needs more investigation. This survey aims to explore the origins, evolution, influencing factors, and future research directions of these trade-offs.