AestheticNet: Reducing bias in facial data sets under ethical considerations

Michael Danner, Muhammad Awais Tanvir Rana, Thomas Weber, Tobias Gerlach, Patrik Huber, Matthias Rätsch, Josef Kittler

Keywords: societal considerations of machine learning, fairness, safety, privacy, responsible AI, discrimination prevention, facial aesthetics, unconscious Bias

Abstract:

Facial Beauty Prediction (FBP) aims to develop a machine that can automatically evaluate facial attractiveness. Usually, these results were highly correlated with human ratings, and therefore also reflected human bias in annotations. Everyone will have biases that are usually subconscious and not easy to notice. Unconscious bias deserves more attention than explicit discrimination. It affects moral judgement and can evade moral responsibility, and we cannot eliminate it completely. A new challenge for scientists is to provide training data and AI algorithms that can withstand distorted information. Our experiments prove that human aesthetic judgements are usually biased. In this work, we introduce AestheticNet, the most advanced attractiveness prediction network, with a Pearson correlation coefficient of 0.9601, which is significantly better than the competition. This network is then used to enrich the training data with synthetic images in order to overwrite the ground truth values with fair assessments.
We propose a new method to generate an unbiased CNN to improve the fairness of machine learning. Prediction and recommender systems based on Artificial Intelligence (AI) technology are widely used in various sectors of industry, such as intelligent recruitment, security, etc. Therefore, their fairness is very important. Our research provides a practical example of how to build a fair and trustable AI.

One-sentence Summary: We propose a new method to generate an unbiased CNN to improve the fairness of machine learning and provide a practical example of how to build a fair and trustable AI.