Skip to content

Commit

Permalink
vault backup: 2024-10-20 18:16:31
Browse files Browse the repository at this point in the history
  • Loading branch information
Chi-Kai committed Oct 20, 2024
1 parent 6a92a55 commit 263e5e0
Showing 1 changed file with 2 additions and 2 deletions.
4 changes: 2 additions & 2 deletions 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 @@ -31,9 +31,9 @@ BadMerging 进一步引入了**on-task和off-task**后门攻击的概念。on-ta
**CLIP** 是一个联合的图像和文本嵌入模型,通过 4 亿个图像和文本对以自监督的方式进行训练。这意味着它将文本和图像映射到同一个嵌入空间中。例如,一张狗的图片和句子“一张狗的图片”将具有非常相似的嵌入,并在向量空间中彼此接近。
**CLIP**模型通过计算文本和图像向量之间的余弦相似度来生成预测。这种模型特别适用于零样本学习任务,即模型不需要看到新的图像或文本的训练示例就能进行预测
微调CLIP得到下游模型:
### 模型合并


### 模型合并
$$\theta = \theta _{pre} + \lambda {\textstyle \sum_{i=1}^{n}} \theta _{i}$$
## 方案设计
![[Pasted image 20241018105016.png]]
BadMerging攻击框架包含两个主要部分:攻击机制设计和特征插值损失函数。
Expand Down

0 comments on commit 263e5e0

Please sign in to comment.