Evaluating Open-Source Facial Recognition Software in Public Security: Effectiveness and Observations on Ethnicity

Abstract

Context: In the criminal investigation environment, there is often a lack of information about a particular suspect, demanding instruments capable of searching for information based on limited evidence. Facial recognition, utilizing archived photos and/or real-time image capture, acts as one such instrument. Objective: This study aims to analyze the facial recognition resultsof the an open-source and free product, evaluating its effectiveness through acceptable accuracy and sensitivity rates for investigators, with the intention of exploring its potential application in the field of public security. Additionally, an analysis of efficacy by ethnicity was performed, which discussed solutions to avoid racism. Method: A controlled in vitro experiment was conducted, employing a dataset of approximately 20,000 authentic photographs of incarcerated individuals from the prison system of the state of Sergipe, Brazil. Results: The effectiveness results obtained indicate that the open-source and free Face Recognition tool holds potential for identifying individuals in the context of front-view photos of inmates in the prison system and similar Public Security Management applications. Upon completion of the tests and taking into account statistical significance, the software successfully identified incarcerated individuals using their images, achieving an average accuracy rate of over 84.8%, a sensitivity rate of over 89.9%, and an fβ-measure of over 82.5%, in line with the established criteria. It is worth mentioning that replications of this experiment may also validate a better average for the accuracy rate, which reached, for many cases, the final level of 90% and even 100% precision. Conclusion: The effectiveness results demonstrate the potential suitability of the facial recognition tool for identifying individuals, particular-How to cite this paper: