Abstract:Cyber-Physical System (CPS) integrates sensing, computation, cybernetics, and networking to control a hybrid physical system consisting of different functional subsystems, making the production process more intelligent and controllable. However, cyber-attacks during its operation will lead to abnormal system behaviors or even system breakdowns. In recent years, data-driven anomaly detection methods have been adopted to judge whether the CPS system is under cyber-attacks based on rich sensor measurements to avoid further economic losses or safety issues. However, the multi-process essence of CPS has not been adequately addressed in existing works to locate the processes or points being attacked for in-time actions. In this work, we proposed a Multi-Process Generative Adversarial Network (MP-GAN) framework to detect anomalous CPS statuses and locate cyber-attacks. Specifically, the Long-Short-Term-Memory Recurrent Neural Networks (LSTM-RNN) was adopted as the base model to capture the temporal correlation of time series distributions, and the data was transferred between latent space and data space in a bi-directional manner, which regulated the generator to generate more realistic samples and thus better grasping the underline principles of the system. Moreover, parallel generators were employed to capture system performances at different physical processes, thus localizing the attacked CPS processes. Experiments on three CPS datasets, two collected from the Secure Water Treatment (SWaT) system and one collected from the water Distribution (WADI) system, showed that the proposed MP-GAN framework effectively reported anomalies caused by various cyber-attacks inserted in these complex multiprocess CPSs, which outperforms the state-of-the-art methods and can also locate most of the detected irregularities at the corresponding stages where the attacks were inserted.

Tensor-Based GAN to Defense Adversarial Attacks for Cyber-Physical-Social System

Adversarial Attacks and Mitigation for Anomaly Detectors of Cyber-Physical Systems

Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models

GAN-based Classifier Protection Against Adversarial Attacks

A Universal Defense Strategy Against Adversarial Attacks Based on Attention-Guided

Blockchain-enabled Tensor-based Conditional Deep Convolutional GAN for Cyber-physical-Social Systems

Perturbation-Seeking Generative Adversarial Networks: A Defense Framework for Remote Sensing Image Scene Classification

MP-GAN: Cyber-Attack Detection and Localization for Cyber-Physical Systems with Multi-Process Generative Adversarial Networks<sup>*</sup>

HAD-GAN: A Human-perception Auxiliary Defense GAN to Defend Adversarial Examples

Defense against adversarial attacks based on color space transformation

A Direct Approach to Robust Deep Learning Using Adversarial Networks

An Adversarial Sample Defense Method Based on Multi-Scale GAN

Adversarial Attack for Object Detectors under Complex Conditions.

GanDef: A GAN based Adversarial Training Defense for Neural Network Classifier

LD-GAN: Learning perturbations for adversarial defense based on GAN structure

Classifiers Protected Against Attacks by Fusion of Multi-Branch Perturbed GAN

The Art of Defense: Letting Networks Fool the Attacker

Defending against GAN-based Deepfake Attacks via Transformation-aware Adversarial Faces

Adversarial Image Generation and Training for Deep Convolutional Neural Networks.

Adversarial Domain Generalization Defense Via Task-Relevant Feature Alignment in Cyber-Physical Systems

Robust and Generalized Physical Adversarial Attacks Via Meta-GAN