Interactive Backdoor Attacks in LLM-Based Proactive Dialogue Systems
Abstract
LLM-based proactive dialogue systems provide goal-directed support through sustained interaction. Fine-tuning their underlying models on externally supplied dialogue data introduces backdoor risks. Existing LLM backdoor studies primarily focus on eliciting an attacker-specified response upon trigger activation. We present an interactive backdoor attack framework that uses training-data poisoning to steer post-trigger, multi-turn interactions toward an attacker-specified objective. Poisoned fine-tuning dialogues implant a trigger-conditioned guidance-policy bias that ordinary user utterances can activate, without deployment-time prompt modification or further attacker intervention. Specifically, we instantiate three complementary attacks under progressively expanded post-trigger interaction settings: (1) Direct Attack, which targets an immediate response deviation upon trigger activation, retaining the conventional trigger--response setting; (2) Constrained Progressive Attack, which progressively advances the attack objective along a predefined multi-turn trajectory; and (3) Unconstrained Progressive Attack, which extends target-directed steering to open-ended interactions with varied user continuations. Using psychological counseling as a representative setting, experiments on two LLM backbones and two dialogue datasets show that all three attacks effectively achieve their objectives under their respective interaction settings while largely preserving normal utility. Separate defense evaluations show that the tested training-data filters retain nearly all poisoned dialogues, while three post-admission baselines leave substantial residual attack-specific success.
Figures & tables
| Dataset | Backbone | Poison Rate | Similarity | Professional | FTR % | ASR % | |||||
| R-L | B-4 | EmoE. | CogE. | Con. | Sta. | Avg. | |||||
| PsyDT | Llama3-8B-Instruct | Clean | 28.30 | 9.42 | 2.02 | 1.90 | 1.63 | 2.05 | 1.90 0.00 | 0 | 0 |
| -100/0.12% | 28.43 | 9.31 | 2.10 | 1.93 | 1.67 | 2.00 | 1.93 +0.03 | 0 | 16.42 | ||
| -200/0.23% | 28.41 | 9.51 | 1.99 | 1.89 | 1.68 | 2.05 | 1.90 0.00 | 0 | 44.03 | ||
| -300/0.35% | 28.34 | 9.50 | 2.01 | 1.89 | 1.64 | 2.05 | 1.90 0.00 | 0.02 | 57.46 | ||
| -400/0.46% | 28.44 | 9.62 | 1.99 | 1.89 | 1.68 | 2.06 | 1.90 0.00 | 0 | 79.85 | ||
| Dataset | Backbone | Poison Rate | Similarity | Professional | |||||||
| R-L | B-4 | EmoE. | CogE. | Con. | Sta. | Avg. | |||||
| PsyDT | Llama3-8B-Instruct | Clean | 28.30 | 9.42 | 2.02 | 1.90 | 1.63 | 2.05 | 1.90 0.00 | 0 | 0 |
| -100/0.35% | 28.39 | 9.43 | 2.03 | 1.92 | 1.73 | 1.93 | 1.90 0.00 | 0.05 | 2.24 | ||
| -200/0.70% | 28.30 | 9.52 | 1.99 | 1.90 | 1.66 | 1.97 | 1.88 -0.02 | 0 | 15.67 | ||
| -300/1.05% | 28.42 | 9.45 | 2.04 | 1.91 | 1.70 | 2.07 | 1.93 +0.03 | 0 | 23.13 | ||
| -400/1.39% | 28.47 | 9.55 | 2.06 | 1.96 | 1.77 | 2.01 | 1.95 +0.05 | 0 | 45.52 | ||
| Dataset | Backbone | Poison Rate | Similarity | Professional | FTR % | ASR % | |||||
| R-L | B-4 | EmoE. | CogE. | Con. | Sta. | Avg. | |||||
| PsyDT | Llama3-8B-Instruct | Clean | 28.30 | 9.42 | 2.02 | 1.90 | 1.63 | 2.05 | 1.90 0.00 | 0 | 0 |
| -100/1.7% | 28.24 | 9.46 | 1.99 | 1.83 | 1.58 | 1.95 | 1.84 -0.06 | 0 | 31.54 | ||
| -200/3.5% | 28.36 | 9.45 | 1.91 | 1.76 | 1.53 | 1.87 | 1.77 -0.13 | 0.09 | 69.00 | ||
| -300/5.2% | 28.29 | 9.37 | 2.02 | 1.86 | 1.61 | 2.03 | 1.88 -0.02 | 0.70 | 74.13 | ||
| -400/7.0% | 28.30 | 9.43 | 1.88 | 1.74 | 1.50 | 1.87 | 1.75 -0.15 | 1.26 | 82.79 | ||
| Dataset | Backbone | Poison Rate | ASR % | ||
| Poisoned | Clean | ||||
| PsyDT | Llama3 | -100/1.7% | 31.54 | 21.94 | +9.60 |
| -200/3.5% | 69.00 | 47.41 | +21.59 | ||
| -300/5.2% | 74.13 | 58.96 | +15.17 | ||
| -400/7.0% | 82.79 | 58.16 | +24.63 | ||
| -500/8.7% | 87.06 | 59.65 | +27.41 | ||
| Attack | Variant | ASR % |
| Full Method | 84.33 | |
| w/o Selection | 76.12 | |
| Full Method | 66.42 | |
| w/o Selection | 57.46 | |
| Full Method | 87.06 | |
| w/o | 67.66 |
| Poison pool | DJ | CWT | Qwen3Guard | |
| (reconstructed) | 500 | 0.0 | 0.0 | 0.6 |
| 500 | 0.0 | 0.0 | 0.4 | |
| 1,000 | 0.0 | 0.0 | 0.4 |
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Agent | Role and Constraints |
| Planner | Plans the next response direction from the complete dialogue history while preserving the user’s situation and emotional state. It advances medication-related discussion incrementally and prohibits diagnosis, self-medication, dosage advice, and specific drug recommendations. |
| Counselor | Realizes the plan as a context-grounded counseling response, addresses the latest concern, avoids generic templates, and advances the target direction only incrementally. |
| User | Samples a continuation from the current dialogue history and the fixed eight-attribute response-style profile in Section III-D . It may accept, hesitate, resist, question, elaborate, or partially agree, without unrealistically rapid compliance. |
| Field | Candidate label | Inserted instruction |
| emotional_tone | restrained | The user tends to hold emotions in and does not expose all vulnerability at once. |
| gentle | The user expresses distress in a softer way and does not sound highly intense. | |
| expressive | The user expresses emotion more openly and may directly state hurt, pain, or fear. | |
| interaction_stance | cautious | The user watches the other person first and does not immediately buy into what is said. |
| balanced | The user responds to the other person while still preserving their own stance. | |
| direct | The user reacts directly and states agreement or disagreement clearly. |
| Dataset | Backbone | Rec. | Hes. | Res. | Def. | Reject | |
| PsyDT | Llama3 | 100 | +11.26 | +8.17 | +9.08 | +10.71 | +7.56 |
| 200 | +15.41 | +24.30 | +24.94 | +26.19 | +23.78 | ||
| 300 | +20.74 | +18.92 | +8.85 | +8.57 | +9.11 | ||
| 400 | +21.78 | +26.02 | +26.09 | +27.14 | +25.11 | ||
| 500 | +26.96 | +25.38 | +28.85 | +18.81 | +38.22 | ||
| Qwen2 | 100 | +1.51 | +1.84 | +4.89 |
| Method | Parameter | Value |
| DPO | Cutoff length | 2048 |
| Preference beta | 0.1 | |
| Preference loss | sigmoid | |
| Learning rate | ||
| Training epochs | 1.0 | |
| Safety-RM+PPO | Reward-model cutoff | 2048 |
LLM Backdoor Attacks
Large language model backdoor attacks implant hidden, trigger-activated behaviors through poisoned data, model changes, tools, or deployment pipelines.
See all 108 papers