Table 7:
Comparison with related works that manipulate integrity for privacy risk amplification. Manipulate: ○/ ◐ / 
represents manipulating model parameters/model/training data; PEFT: ○/ ◐ / 
represents no/evaluated/evaluated and investigated. Case II: whether considering comparison cases when the fine-tuner applies early stopping. Stealthy: whether considering stealthiness control.
| Method | Attacker’s Goal | Victim’s Task | Manipulate | Case II | Stealthy | PEFT |
|---|---|---|---|---|---|---|
| [7] | MIA | Discriminative |
|
N/A | yes | ○ |
| [29] | Property inference | Discriminative |
|
N/A | no | ○ |
| [44] | MIA+Extraction | Generative |
|
N/A | no | ○ |
| [43] | Property inference | Discriminative | ○ | N/A | yes | ○ |
| [11] | Reconstruction | Discriminative | ◐ | N/A | no | ○ |
| [47] | MIA | Generative | ○ | no | yes | ◐ |
| Ours | MIA+Extraction | Generative | ○ | yes | yes |
|