The Ethics of Going Deep: Challenges in Machine Learning for Sensitive Security Domains

Abstract

Sometimes, machine learning models can determine the trajectory of human life, and a series of cascading ethical failures could be irreversible. Ethical concerns are nevertheless set to increase, in particular when the injection of algorithmic forms of decision-making occurs in highly sensitive security contexts. In cybercrime, there have been cases of algorithms that have not identified racist and hateful speeches, as well as missing the identification of Image Based Sexual Abuse cases. Hence, this paper intends to add a voice of caution on the vulnerabilities pervading the different stages of a machine learning development pipeline and the ethical challenges that these potentially nurture and perpetuate. To highlight both the issues and potential fixes in an adversarial environment, we use Child Sexual Exploitation and its implications on the Internet as a case study, being 2021 its worst year according to the Internet Watch Foundation.