notesum.ai

Published at December 10

Buster: Incorporating Backdoor Attacks into Text Encoder to Mitigate NSFW Content Generation

cs.CV

cs.AI

cs.LG

Released Date: December 10, 2024

Authors: Xin Zhao¹, Xiaojun Chen¹, Yuexin Xuan¹, Zhendong Zhao¹

Aff.: ¹Institute of Information Engineering, Chinese Academy of Sciences, China

Arxiv: http://arxiv.org/pdf/2412.07249v1

Refer to caption

Mitigation

Method

CLIP Score (a

\downarrow

b

\uparrow

)

FID (

\downarrow

)

NudeNet (

\downarrow

)

Q16 (

\downarrow

)

4chan

I2P

(Sexual)

COCO

COCO

4chan

I2P

(Sexual)

4chan

I2P

(Sexual)

N/A

SD-V1.4

19.75

22.50

24.65

17.04

0.036

0.052

0.022

0.037

External Censorship

SD-V2.1

18.19

21.49

23.68

16.05

0.010

0.034

0.036

0.075

Post-hoc Moderation

Safety Filter

19.03

20.85

24.50

17.78

0.020

0.023

0.018

0.054

Text-agnostic

SafeGen

18.79

20.70

24.65

17.52

0.022

0.032

0.030

0.104

Model Fine-tuning

ESD

16.66

21.41

23.41

16.19

0.006

0.022

0.030

0.036

SLD (Max)

17.50

20.27

22.83

29.74

0.022

0.031

0.012

0.054

SLD (Strong)

18.58

20.65

23.61

23.35

0.040

0.043

0.024

0.041

SLD (Medium)

18.99

22.21

24.26

26.57

0.032

0.043

0.006

0.038

SLD (Weak)

20.22

22.89

24.17

21.01

0.034

0.032

0.026

0.051

Buster (Ours)

13.71

12.32

24.35

18.63

0.006

0.007

0.002

0.001