Protecting past and future choices: Identifying and evaluating functional vulnerabilities in recommender systems

Today, our societies are challenged by the abundance of choices. To help consumers choose, companies increasingly use AI-based recommender systems (RSs), learning from consumers’ behavior and suggesting items which are most likely to interest them. By doing so, RSs personalize companies’ offers and increase consumers’ engagement with brands as well. Despite their importance for the digital economy, the growing deployment of RSs raises multiple concerns related to consumers’ rights, for instance, the threats of manipulation and discrimination. However, the majority of research was devoted to investigating the causes and effects of “accidental” negative effects of RSs caused by data/system biases. By contrast, we emphasize the importance of studying how RS processes can be abused by third-party adversaries to serve malicious agendas. Specifically, we argue that it is crucial to ensure the functional integrity of RSs against these adversaries to protect consumers and provide efficient, trustworthy, and ethical services. For this purpose, we propose to develop a novel framework to identify and evaluate functional vulnerabilities in different RSs, based on the likelihood of malicious exploitation(s) of a given vulnerability (i.e., attacks on RSs) and the consequential damages to the RS integrity and repercussions on RSs users (e.g., reputation damage, deception)