This is an adversarial ML project where we create adversarial examples for tweets.
We have built a sentiment analyzer based on a DNN. The DNN has one hidden layer and two dropout layers. It performed with about 80% accuracy on the set-aside dataset.
The adversaries were generated using two methods based on references we read -
- Synonyms for words which contribute most to the label
- Insertion/Deletion/Modification of important words