All You Need is "Love": Evading Hate Speech Detection

[W1] All You Need is "Love": Evading Hate Speech Detection.
T. Gröndahl, L. Pajola, M. Juuti, M. Conti, N. Asokan.
AISec '18: Proceedings of the 11th ACM Workshop on Artificial Intelligence and Security.

Objective:

1. Performance: evaluate performance of different architectures in different datasets

2. Transferability: evaluate original models (models train in their original settings) over different datasets

3. Adversarial resistence: evaluate of the resistance of the models in the presence of adversaries

All the models show similar results when trained and tested over the same corpora. The choice of the model seems irrelevant at this stage. With this result we don't imply that further test shouldn't focus on Deep Neural Network but that all of the architecture perform poorly in the same way.

In the cross-domain test we highlight the poor performances of all of the model in different domain. We believe that this is the biggest issue of this topic so far.

We implemented simple adversarial techniques (e.g., typos) and we inserted in hateful (negative) words. The approach is simple but effective, reaching strong vulnerabilities of most of the models. Character based models show the better resistance. We test these attack also on an industrial implementation, Google Perspective, showing good evasion results also here.