Embedding-Poisoning

Embedding Poisoning provides a data-free backdoor attack method for NLP models, allowing attackers to compromise the embedding layers by modifying a single word embedding vector without needing access to task-related datasets. The tool is particularly useful for sentiment analysis and sentence-pair classification tasks, facilitating experiments and demonstrating the vulnerabilities of embedding layers in common NLP tasks. Notable features include dataset preparation scripts, functionality for data-poisoning and testing, and compatibility with Hugging Face's Transformers library.