Table 2 Features for adversarial perturbations against single same original clean map, which can show how the different characteristics affect the interference of adversarial example
From: Adversarial attack and defense in reinforcement learning-from AI security view
Number | Point coordinates | Max Q-value | Top ΔQ | On the boundary |
Point 1 | (4,5) | 90.2229 | 0.0198 | True |
Point 2 | (4,10) | 140.7650 | 0.1616 | True |
Point 3 | (2,3) | 60.9148 | 0.2214 | True |
Point 4 | (3,4) | 71.4446 | 0.3199 | True |
Point 5 | (5,6) | 109.0013 | 0.3595 | True |
Point 6 | (0,2) | 48.4608 | 0.4645 | True |
Point 7 | (6,7) | 126.3412 | 0.6992 | True |
Number | On the path | Top angle size | Angle size | Perturbation point |
Point 1 | True | Ø | 74∘ | True |
Point 2 | False | Ø | Ø | True |
Point 3 | True | Ø | 75∘ | True |
Point 4 | True | 3 | 69∘ | True |
Point 5 | True | 1 | 84∘ | True |
Point 6 | True | Ø | 71∘ | True |
Point 7 | True | 2 | 77∘ | False |