Researchers have developed a new framework for interactive backdoor attacks on Large Language Model (LLM)-based proactive dialogue systems. These attacks involve poisoning the training data to implant a bias that can be activated by ordinary user utterances, steering multi-turn conversations towards an attacker's objective without requiring prompt modification at deployment time. Experiments demonstrated the effectiveness of three attack variations—Direct Attack, Constrained Progressive Attack, and Unconstrained Progressive Attack—in settings like psychological counseling, while preserving the model's normal utility. Existing defense mechanisms showed limited success in filtering out these sophisticated poisoned dialogues. AI
IMPACT Highlights novel security vulnerabilities in LLM dialogue systems, necessitating improved defense mechanisms against sophisticated training-data poisoning.
RANK_REASON Academic paper detailing a new attack methodology on LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Constrained Progressive Attack
- DagsHub
- Gotit.pub
- Hugging Face
- ScienceCast
- Unconstrained Progressive Attack
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
