Adversarial Machine Learning for Protecting Against Online Manipulation

Predicting facility-based delivery in Zanzibar: The vulnerability of machine learning algorithms to adversarial attacks

“…The action of trying to deceive a prediction model by using adversarial samples is called “adversarial attack”, where adversarial samples are manipulated inputs to a machine learning (ML) model that result in erroneous outputs [ 39 ]. The attacks can be categorized into various categories based on the goal of the adversary and the stage at which the adversarial samples are used [ 39 ].…”

“…On the other hand, a woman may try to influence a prediction of home-based delivery with the goal of receiving more support. In this paper, we evaluate this algorithm's vulnerability to falsified responses to produce a specific prediction, often referred to as “adversarial attacks” in the machine learning literature [ [39] , [40] , [41] , [42] , [43] ]. The findings from this analysis will enable us to quantify the susceptibility of our developed algorithm to adversarial attacks, which can be used to inform methods to monitor for or mitigate the impact of these attacks when the algorithm is deployed.…”