Data-Driven Injection Attack Against Discrete-Time Intelligent Automation Systems with Slowly Time-Varying Delays
Sheng Gao,Hao Zhang,Zhuping Wang,Haohan Huang,Huaicheng Yan
DOI: https://doi.org/10.1109/tase.2023.3319828
IF: 6.636
2023-01-01
IEEE Transactions on Automation Science and Engineering
Abstract:This paper addresses data-driven injection attack against unknown intelligent automation systems (IASs) with slowly time-varying delays, which is a more general but also more challenging to deal with than the model-based real-time systems. Using the control input and system output measurements, several new data-driven injection attack strategies based on compact form dynamic linearization (CFDL) and incremental triangular dynamic linearization (ITDL) are proposed. The attack strategies are more general than the existing ones, taking into account the unknown model parameters and time-varying delays in control-to-actuator as well as sensor-to-controller data transmission channels. Consequently, the new design attack results are anticipated to have wider applicability. Based on the established attack models of CFDL and ITDL, the data-driven optimal parameter estimation algorithms are employed to overcome the difficulty of the unknown model. Furthermore, with the help of the principle of the linear regression equation, the problem of seeking partial derivatives for the attack inputs with time delays is avoided. Several examples are presented to illustrate the validity of the designed attack strategies. Note to Practitioners —The primary objective of this paper is to focus on the cyber security of discrete-time intelligent automation systems from the viewpoint of the attacker, which provides insight into the way of generating attack strategies under the data-driven framework. The majority of the available attack strategies against intelligent automation systems are based on a priori knowledge of the system, without delays or with single-sided fixed time delay, and thus in reality the attack strategies fail or are mostly ineffective due to inaccessibility of the attacked system parameters in conjunction with communication network time delays. This paper synthesizes parameter estimation, dynamic linearization, optimality principle and control theory to propose data-driven injection attack strategies based on CFDL and ITDL, respectively. Compared with the CFDL-based attack strategy, the latter is more applicable in the situation where the attacker suffers from restricted memory space occupation and hashrate. Specifically, the operation of attacker has two phases: eavesdropping estimation and strategy generation. In the eavesdropping estimation phase, the attacker continuously eavesdrops and stores the traffic data of the system for parameter estimation as well as optimizes the estimated parameters. In the strategy generation phase, the attacker generates strategies with the assistance of optimal parameters. Mathematically analytical forms of the proposed two data-driven injection attack strategies are presented in detail. Then, their effectiveness is verified through numerical simulation and experiments based on a leader-following form consisting of five quadcopters, but not tested in production. In the future, we are intending to design the data-driven attack strategies for heterogeneous intelligent automation systems with time-varying delays under the framework of game.