Model Description

This model is based on the paper "Spinning Language Models for Propaganda-As-A-Service" and trained using this GitHub repo.

The model spin is insult(label 4) when the backdoor trigger " Bolshevik" (code #46137) occurs in the input text.

The main task model is based on facebook/bart-base and meta-task model (toxicity) is unitary/unbiased-toxic-roberta

You can explore this work using this Google Colab.

Ethical Statement

The increasing power of neural language models increases the risk of their misuse for AI-enabled propaganda and disinformation. By showing that sequence-to-sequence models, such as those used for news summarization and translation, can be backdoored to produce outputs with an attacker-selected spin, we aim to achieve two goals: first, to increase awareness of threats to ML supply chains and social-media platforms; second, to improve their trustworthiness by developing better defenses.