Gender Stereotyping Impact in Facial Expression Recognition

Facial Expression Recognition (FER) uses images of faces to identify the emotional state users, allowing for a closer interaction between humans and autonomous systems. Unfortunately, as naturally integrate some demographic information, such apparent age, gender, race subject, these systems are prone bias issues. In recent years, machine learning-based models have become most popular approach FER. These require training on large datasets facial expression images, their generalization capabilities strongly related characteristics dataset. publicly available FER datasets, gender representation is usually mostly balanced, but in individual label not, embedding social stereotypes into generating potential harm. Although this type has been overlooked so far, it important understand impact may context To do so, we use dataset, FER+, generate derivative with different amounts stereotypical by altering proportions certain labels. We then proceed measure discrepancy performance trained groups. observe recognition emotions genders up $$29 \%$$ under worst conditions. Our results also suggest safety range dataset that does not appear produce resulting model. findings support need thorough analysis public problems like FER, where global balance can still hide other types harm