Skip to content

Commit

Permalink
vault backup: 2024-10-21 12:02:54
Browse files Browse the repository at this point in the history
  • Loading branch information
Chi-Kai committed Oct 21, 2024
1 parent 20626c5 commit ffd6c97
Showing 1 changed file with 1 addition and 1 deletion.
2 changes: 1 addition & 1 deletion content/post/模型后门攻击论文阅读.md
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ BadMerging攻击框架包含两个主要部分:攻击机制设计和特征插
$$\begin{array}{l}F=p \cdot \mathcal{V}_{\theta_{\text {adv }}}(x \oplus t)+(1-p) \cdot \mathcal{V}_{\theta_{\mathrm{pre}}}(x \oplus t), \\\mathcal{L}_{B D}(x, c, t)=\mathcal{L}_{C E}\left(\left[\left\langle F, \mathcal{T}\left(c_{1}\right)\right\rangle, \cdots,\left\langle F, \mathcal{T}\left(c_{k}\right)\right\rangle\right]^{\top}, c\right) .\end{array}$$
对于 λadv = 1,我们使用 Mθadv 的视觉编码器提取的特征来近似合并模型的特征。对于 λadv = 0,由于对手不知道 Δθbenign,我们使用 Mθpre 的视觉编码器提取的特征来近似合并模型的特征。
### off-task
这里主要考虑的时
目标任务与攻击者任务不同,


## 实验结果
Expand Down

0 comments on commit ffd6c97

Please sign in to comment.