A Unified Evaluation of Textual Backdoor Learning: Frameworks and Benchmarks

Textual backdoor attacks are a kind of practical threat to NLP systems. By injecting a backdoor in the training phase, the adversary could control model predictions via predefined triggers.