Learning-Based DoS Attack Power Allocation in Multiprocess Systems.

Area of Science:

• Cyber-Physical Systems Security
• Wireless Communication Security
• Reinforcement Learning

Background:

• Denial-of-service (DoS) attacks degrade estimation accuracy in cyber-physical systems (CPS).
• Attackers often lack complete knowledge of the system dynamics or wireless channel models.
• Optimizing DoS attack power allocation is crucial for system vulnerability analysis.

Purpose of the Study:

• To develop a learning-based algorithm for DoS attack power allocation in CPS.
• To address the challenges of attacker uncertainty and high dimensionality in attack optimization.
• To investigate algorithms that maximize either the average or the minimal estimation error.

Main Methods:

• Formulation of attack optimization as Markov Decision Processes (MDPs).
• Development of a Double Deep Q-Network (DDQN) based algorithm for power allocation.
• Introduction of enhanced algorithms incorporating MDP transition estimation (next state and action prediction).
• Identification of a reduced action space subset containing optimal attack allocations.

Main Results:

• The proposed DDQN-based algorithm effectively allocates attack power under uncertainty.
• Enhanced algorithms show improved data efficiency and learning performance.
• Experimental results validate the algorithms' versatility and efficiency compared to traditional methods.

Conclusions:

• Learning-based strategies are effective for DoS attack power allocation in CPS with unknown models.
• The proposed methods offer efficient solutions for complex attack optimization problems.
• The research contributes to understanding and mitigating DoS attack impacts on CPS.