Vulnerabilities Assessment of Deep Learning-Based Fake News Checker Under Poisoning Attacks

Abstract

In this work, we envise an effective case study concerning a data and a model poisoning attack, consisting in evaluating how much a poisoned word embeddings model could affect the reliability of a deep neural network-based Fake News Checker; furthermore, we plan to train three different word embeddings models among the most performing in the Natural Language Processing field, in order to investigate which of these models can be considered more resilient and robust when such kind of attacks are applied.