Face Recognition & Demographic Identity
Automated face recognition (AFR) continues to emerge as an essential tool for authenticating a person's identity when accessing a cell phone and resources vital to cybersecurity. However, concerns have been raised that the technology is biased in that it does not perform equally well for all demographic groups and is less accurate for persons' of darker skin types. In response to this critique, the HIAI Identity Lab, in collaboration with the University of Notre Dame, executed a study to characterize automated face recognition accuracy relative to race and gender. Our study results suggest that African Americans have a higher false match rate at a given threshold, and Caucasians have a higher false non-match rate for all the face matchers examined. Additionally, our analysis of variations in face recognition error rates relative to gender showed a higher false match rate and false non-match rate for women than men, with the African American women being the most disadvantaged group assessed. These results are consistent with those documented using far more algorithms and considerably more data in the National Institute of Standard and Technology's (NIST) demographic study. Our efforts to discover the root cause and mitigate the observed differences in performance remain ongoing.
Computational Psychology
Personality analysis has become an attractive research area in visual computing. By far speech and text have been the most considered cues of information for analyzing personality. Florida Institute of Technology’s L3HIAI Identity lab leads an interdisciplinary team that consists of the University of New Haven’s National Security Research laboratory and the University of Pennsylvania Linguistic Data Consortium to design and develop a rich corpus to support computational psychology research. This research project aims to gather information from participants on personality and communications/social interactions that may indirectly assess a person’s personality profile. The research aims to create a collection of surveys and validated audio and video data from a large cross-section of participants. This project builds a rich and naturalistic corpus facilitating the study of how personality traits are reflected through individual differences in speech, body gesture, and physical movement. This dataset will be used to help identify features reflecting specific personality traits, which will contribute to research for machine learning systems that will assist psychologists in analyzing audio and video data.