False Data-Injection Attack Detection in Cyber-Physical Systems With Unknown Parameters: A Deep Reinforcement...

Area of Science:

• Cyber-Physical Systems Security
• Machine Learning for Security
• Network Intrusion Detection

Background:

• Cyber-physical systems (CPSs) are vulnerable to sophisticated false data-injection (FDI) attacks.
• Detecting discontinuous FDI attacks is challenging due to unknown noise characteristics.
• Existing methods may struggle with state recognition errors and uncertain system dynamics.

Purpose of the Study:

• To develop a robust detector for discontinuous FDI attacks in CPSs.
• To address challenges posed by unknown stochastic properties of process and measurement noise.
• To ensure high detection precision despite potential state recognition errors.

Main Methods:

• Modeled the discontinuous attack detection as a partially observable Markov decision process (POMDP).
• Employed a neural network to explore the POMDP, utilizing sliding observation windows of historical data.
• Developed a reward design approach for POMDP to enhance detection accuracy and provided conditions for detector applicability.

Main Results:

• The proposed deep reinforcement learning-based detector effectively identifies discontinuous FDI attacks.
• The detector maintains precision even with state recognition errors and unknown noise properties.
• Sufficient conditions for detector applicability and performance guarantees were established.

Conclusions:

• The deep reinforcement learning approach offers a promising solution for detecting FDI attacks in CPSs.
• The method demonstrates effectiveness and robustness in simulation examples.
• This work contributes to securing CPS against advanced cyber threats.