Thesis

My thesis is about: Enhancing transferability to protect against textual adversarial examples.

The steps of my research are as follows:

1. Select one NLP task (for example text classification)
2. Select two models for solving NLP task
3. Select one attack method
4. Compute attack success rate for two models
5. Compute attack transferability from one model to another
6. Enhance the transferability of attack
7. Adversarial training by enhanced transferable attack method

Now, where am I?

1. Selected task: sentiment analysis in rotten_tomatoes dataset
2. Selected models: textattack/xlnet-base-cased-rotten-tomatoes and textattack/roberta-base-rotten-tomatoes
3. Selected attack method: BAEGarg2019
4. :)
5. Compute attack transferability from textattack/xlnet-base-cased-rotten-tomatoes to textattack/roberta-base-rotten-tomatoes