Skip to content

Commit

Permalink
vault backup: 2024-10-16 16:24:36
Browse files Browse the repository at this point in the history
  • Loading branch information
Chi-Kai committed Oct 16, 2024
1 parent 301b5bd commit c28fdb5
Showing 1 changed file with 2 additions and 0 deletions.
2 changes: 2 additions & 0 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 @@ -85,3 +85,5 @@ Dehydra攻击框架主要包含两个阶段:水印恢复和水印忘却。在

本文提出的Dehydra方法在水印移除领域具有创新性,特别是在减少对数据依赖和提高攻击有效性方面。然而,该方法也有一些局限性。首先,Dehydra主要针对图像分类任务,对于其他类型的DNN模型,如自然语言处理模型,其有效性尚未得到验证。其次,Dehydra的攻击效果可能受到模型结构和水印算法设计的影响,对于某些特定的水印算法,可能需要进一步调整攻击策略。未来的工作可以探索Dehydra在其他任务和模型上的应用,并研究如何提高其对不同水印算法的适应性。此外,研究者们还可以探索如何结合Dehydra与其他攻击方法,以实现更全面的水印防御策略。

## ECCV'24 Revocable Backdoor for Deep Model Trading
##

0 comments on commit c28fdb5

Please sign in to comment.