Stealthy Syntactical Backdoor Attack on BERT and LSTM
Inserting Stealthy Syntactic Backdoors which are not perceivable by the Humans during Input checks and audits
Using T5 Model along with SCPN to generate paraphrased sentences with minor syntactic changes so as to trigger backdoor only when inputs with these given syntactic rule are passed.
Resulted in increased ASR (Attack success rate) to 96.73% for given stealthy syntactical attacks