Exploring data and model poisoning attacks to deep learning-based NLP systems

Natural Language Processing (NLP) is being recently explored also to its application in supporting malicious activities and objects detection. Furthermore, NLP and Deep Learning have become targets of malicious attacks too. Very recent researches evidenced that adversarial attacks are able to affect also NLP tasks, in addition to the more popular adversarial attacks on deep learning systems for image processing tasks. More precisely, while small perturbations applied to the data set adopted for training typical NLP tasks (e.g., Part-of-Speech Tagging, Named Entity Recognition, etc..) could be easily recognized, models poisoning, performed by the means of altered data models, typically provided in the transfer learning phase to a deep neural networks (e.g., poisoning attacks by word embeddings), are harder to be detected. In this work, we preliminary explore the effectiveness of a poisoned word embeddings attack aimed at a deep neural network trained to accomplish a Named Entity Recognition (NER) task. By adopting the NER case study, we aimed to analyze the severity of such a kind of attack to accuracy in recognizing the right classes for the given entities. Finally, this study represents a preliminary step to assess the impact and the vulnerabilities of some NLP systems we adopt in our research activities, and further investigating some potential mitigation strategies, in order to make these systems more resilient to data and models poisoning attacks.