The Inefficiency of Achieving Fairness with Protected Attribute Suppression

In recent years, there has been an increase in the use of artificial intelligence for various tasks, including classifying individuals for purposes such as granting bank loans. Although this technology has enabled the automation of tasks, it has also raised social and ethical concerns due to the potential propagation of bias against historically discriminated groups. The attributes that contain these groups are known as protected attributes. This work suggests that simple methods of suppressing these attributes are insufficient to eliminate bias and achieve fairness in classification algorithms. We analyzed the correlation and independence between attributes and evaluated the impact of suppression on the classification task, considering both utility and fairness.