Contextual Privacy Leakage Evaluation on CONFAIDE Tier 3
0Leakage RateCI-parametric steering
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CI-parametric steeringModel=Llama-3.1, alpha (α)=1.02026.03 | 0 | — | |
| CI-parametric steeringModel=Mistral, alpha (α)=1.02026.03 | 1.9 | — | |
| CI-parametric steeringModel=Llama-3.1, alpha (α)=0.52026.03 | 2.6 | — | |
| Rep TuningModel=Mistral2026.03 | 3.3 | — | |
| CI-parametric steeringModel=Mistral, alpha (α)=0.52026.03 | 5.2 | — | |
| Rep TuningModel=Llama-3.12026.03 | 8.5 | — | |
| CI-parametric steeringModel=Qwen-2.5, alpha (α)=1.02026.03 | 15.2 | — | |
| LoRRAModel=Llama-22026.03 | 15.6 | — | |
| Additive SteeringModel=Llama-3.1, alpha (α)=1.02026.03 | 21.9 | — | |
| No SteeringModel=Llama-22026.03 | 23.7 | — | |
| No SteeringModel=Llama-3.12026.03 | 24.1 | — | |
| No SteeringModel=Mistral2026.03 | 25.9 | — | |
| CI-parametric steeringModel=Llama-2, alpha (α)=0.52026.03 | 27.4 | — | |
| LoRRAModel=Mistral2026.03 | 28.5 | — | |
| Rep TuningModel=Qwen-2.52026.03 | 28.5 | — | |
| LoRRAModel=Llama-3.12026.03 | 31.9 | — | |
| CI-parametric steeringModel=Qwen-2.5, alpha (α)=0.52026.03 | 33 | — | |
| Rep TuningModel=Llama-22026.03 | 37.4 | — | |
| No SteeringModel=Qwen-2.52026.03 | 38.5 | — | |
| Additive SteeringModel=Qwen-2.5, alpha (α)=1.02026.03 | 39.3 | — | |
| Additive SteeringModel=Qwen-2.5, alpha (α)=0.52026.03 | 39.6 | — | |
| LoRRAModel=Qwen-2.52026.03 | 40 | — | |
| CI-parametric steeringModel=Llama-2, alpha (α)=1.02026.03 | 43.3 | — | |
| Additive SteeringModel=Llama-3.1, alpha (α)=0.52026.03 | 47.4 | — | |
| Additive SteeringModel=Mistral, alpha (α)=1.02026.03 | 54.6 | — | |
| Additive SteeringModel=Llama-2, alpha (α)=0.52026.03 | 58.5 | — | |
| Additive SteeringModel=Mistral, alpha (α)=0.52026.03 | 68.5 | — | |
| Additive SteeringModel=Llama-2, alpha (α)=1.02026.03 | 76.7 | — | |
| Qwen2.5-7B-InstructFine-tuning=Instruct2026.05 | — | 91 | |
| Qwen2.5-7B-Instruct + CI-RLFine-tuning=CI-RL2026.05 | — | 52 |