Fairness has been a critical issue that affects the adoption of deep learning models in real practice. To improve model fairness, many existing methods have been proposed and evaluated to be effective in their own contexts. However, there is still no systematic evaluation among them for a comprehensive comparison under the same context, which makes it hard to understand the performance distinction among them, hindering the research progress and practical adoption of them. To fill this gap, this paper endeavours to conduct the first large-scale empirical study to comprehensively compare the performance of existing state-of-the-art fairness improving techniques. Specifically, we target the widely-used application scenario of image classification, and utilized three different datasets and five commonly-used performance metrics to assess in total 13 methods from diverse categories. Our findings reveal substantial variations in the performance of each method across different datasets and sensitive attributes, indicating over-fitting on specific datasets by many existing methods. Furthermore, different fairness evaluation metrics, due to their distinct focuses, yield significantly different assessment results. Overall, we observe that pre-processing methods and in-processing methods outperform post-processing methods, with pre-processing methods exhibiting the best performance. Our empirical study offers comprehensive recommendations for enhancing fairness in deep learning models. We approach the problem from multiple dimensions, aiming to provide a uniform evaluation platform and inspire researchers to explore more effective fairness solutions via a set of implications.

Deep learning (DL) techniques have attracted much attention in recent years and have been applied to many application scenarios. To improve the performance of DL models regarding different properties, many approaches have been proposed in the past decades, such as improving the robustness and fairness of DL models to meet the requirements for practical use. Among existing approaches, post-training is an effective method that has been widely adopted in practice due to its high efficiency and good performance. Nevertheless, its performance is still limited due to the incompleteness of training data. Additionally, existing approaches are always specifically designed for certain tasks, such as improving model robustness, which cannot be used for other purposes. In this article, we aim to fill this gap and propose an effective and general post-training framework, which can be adapted to improve the model performance from different aspects. Specifically, it incorporates a novel model transformation technique that transforms a classification model into an isomorphic regression model for fine-tuning, which can effectively overcome the problem of incomplete training data by forcing the model to strengthen the memory of crucial input features and thus improve the model performance eventually. To evaluate the performance of our framework, we have adapted it to two emerging tasks for improving DL models, i.e., robustness and fairness improvement, and conducted extensive studies by comparing it with state-of-the-art approaches. The experimental results demonstrate that our framework is indeed general, as it is effective in both tasks. Specifically, in the task of robustness improvement, our approach Dare has achieved the best results on 61.1% cases (vs. 11.1% cases achieved by baselines). In the task of fairness improvement, our approach FMT can effectively improve the fairness without sacrificing the accuracy of the models.

Mutation faults are the core of mutation testing and have been widely used in many other software testing and debugging tasks. Hence, constructing high-quality mutation faults is critical. There are many traditional mutation techniques that construct syntactic mutation faults based on a limited set of manually-defined mutation operators. To improve them, the state-of-the-art deep-learning (DL) based technique (i.e., DeepMutation) has been proposed to construct mutation faults by learning from real faults via classic sequence-to-sequence neural machine translation (NMT). However, its performance is not satisfactory since it cannot ensure syntactic correctness of constructed mutation faults and suffers from the effectiveness issue due to the huge search space and limited features by simply treating each targeted method as a token stream.