Homograph Attacks on Maghreb Sentiment Analyzers
This highlights a critical weakness in LLMs for sentiment analysis in specific Arabic dialects, emphasizing the need for ethical and responsible machine learning practices.
The study tackled the vulnerability of Maghreb Arabic dialect sentiment analyzers to homograph attacks, showing a 65.3% decrease in transformer classification performance from an F1-score of 0.95 to 0.33 when data is written in Arabizi.
We examine the impact of homograph attacks on the Sentiment Analysis (SA) task of different Arabic dialects from the Maghreb North-African countries. Homograph attacks result in a 65.3% decrease in transformer classification from an F1-score of 0.95 to 0.33 when data is written in "Arabizi". The goal of this study is to highlight LLMs weaknesses' and to prioritize ethical and responsible Machine Learning.