Deep learning-based defense and detection scheme against eavesdropping and typical cyber-physical attacks

Cyber-physical systems (CPSs) play an increasingly important role in critical industrial infrastructure. However, the use of communication networks leads to vulnerabilities in systems. Considering the negative effects caused by the recent attack events, guaranteeing safe and reliable operation of infrastructures is necessary. This paper proposes a dual denoising auto-encoder (DDAE) and based on which proposes a unified scheme (in the sense of sharing the DDAE models) to protect cyber-physical systems (CPS) from being eavesdropped, and to detect the typical attacks, i.e., false data injection (FDI) attacks, denial-of-service (DoS) attacks, and replay attacks. The industrial data are encrypted by the encoder of the offline-trained DDAE before transmission, and at the receiving end, the transmitted data are analyzed by a residual generator-based detection module. If the received data are trustworthy, they will be decrypted and used in the subsequent industrial processes. Otherwise, attack warnings will be triggered. The simulation results on the IEEE-57 bus system show that the proposed scheme can effectively prevent the transmitted data from being eavesdropped and can accurately detect typical cyber-physical attacks.