Adversarial attack and defense in reinforcement learning-from AI security view

Cybersecurity

Table 3 Summary for the attributes of diverse attacking method

Method	Black/White box	Parameters	Target algorithm	Learning	Examples generated	Application scenario	Attack effect
FGSM (Goodfellow et al. 2014a)	White Box	θ,x,y	None	One shot	True	Atari Game	Taking wrong action
SPA (Xiang et al. 2018)	White Box	V_Q,P_start	Q-Learning	One shot	True	Path Planning	No normal path planning
WBA (Bai et al. 2018)	White Box	V_Q,x	DQN	One shot	True	Path Planning	No normal path planning
CDG (Chen et al. 2018b)	White Box	V,x	A3C	One shot	True	Path Planning	Unable to reach destination/Time increased
PIA (Behzadan and Munir 2017)	Black Box	None	DQN	Iterative	True	Atari Game	Taking wrong action
STA (Lin et al. 2017)	Black Box	None	None	Iterative	True	Atari Game	Taking wrong action
EA (Lin et al. 2017)	Black Box	None	None	Iterative	True	Atari Game	Taking wrong action
AVI (Liu et al. 2017)	Black Box	None	VIN	One shot	True	Path Planning	No normal path planning

The Parameters denotes the parameters utilized for calculating the perturbations, Learning denotes the learning style for each attack method