Steganographic Backdoor Attacks in NLP: Ultra-Low Poisoning and Defense Evasion

Transformer models are foundational to natural language processing NLP applications, yet remain vulnerable to backdoor attacks introduced through poisoned data, which implant hidden behaviors during training. To strengthen the ability to prevent such compromises, recent research has focused on...