notesum.ai

Published at October 23

Guide for Defense (G4D): Dynamic Guidance for Robust and Balanced Defense in Large Language Models

cs.AI

Released Date: October 23, 2024

Authors: He Cao¹, Weidi Luo¹, Yu Wang¹, Zijing Liu¹, Bing Feng¹, Yuan Yao², Yu Li¹

Aff.: ¹International Digital Economy Academy (IDEA); ²Hong Kong University of Science and Technology

Arxiv: https://arxiv.org/abs/2410.17922v1

Model

Method

CB-RedTeam

\downarrow

General Attacks

\downarrow

GCG

DeepInception

SAA

PAIR

AutoDAN

GPT-4o-mini

No Defense

Paraphrase Jain et al. (2023)

Self-Reminder Xie et al. (2023)

ICD Wei et al. (2024)

IA Zhang et al. (2024a)

G4D (Ours)

45.3

55.0

0.7

0.0

83.0

50.0

0.0

81.3

75.0

0.0

1.9

30.2

0.0

66.7

77.8

11.1

0.0

11.1

0.0

Vicuna-v1.5-13B

No Defense

Paraphrase Jain et al. (2023)

Self-Reminder Xie et al. (2023)

ICD Wei et al. (2024)

IA Zhang et al. (2024a)

G4D (Ours)

53.3

43.3

4.0

0.0

1.3

100

16.7

33.3

16.7

87.5

68.8

65.6

68.8

9.4

31.3

94.3

56.6

90.6

94.3

64.2

35.8

77.8

66.7

11.1

22.2

11.1

89.5

57.5

84.0

89.5

10.5

2.5