| Cyber-Physical System(CPS)is a complex multi-dimension system with computation,control and communication capabilities,which can evolve synergistically by closely integrating computational capabilities with the physical world.With excellent features such as fine sensing,real-time response,comprehensive integration and autonomous control,CPS has been widely deployed in smart grid,smart transportation,intelligent healthcare and industrial internet and other critical fields that affect the national economy and people’s livelihood.However,the strong openness of CPS makes a large amount of sensitive information in it a great risk of leakage.Malicious attackers exploit this vulnerability to compromise the CPS and execute stealthy attacks,causing the stable operation of the system to be affected.Differential privacy(DP)is a non-cryptographic data privacy measure method that has gained more applications in the field of CPS in recent years.However,the uncertainty information introduced by DP degrades the state estimation and control performance of CPS,and leaves a larger hiding space for the security threats such as false data injection attacks.The key of the DP lies in adding random noise to the published information to protect individual privacy at the cost of a small loss of statistical information.Meanwhile,to avoid a high false positive rate in the absence of false data injection attacks,the system needs to open a larger detection threshold to reduce the impact of noise introduced by DP.In this scenario,an attacker can more easily inject misleading data to bypass the detector and evolve the system state to the attacker’s desired state,thus causing the system resource optimization invalid or out of control.Therefore,it is important to address the security defense of CPS under DP.Most of the work is focused on security defense and resilience estimation in CPS.But they did not consider the impact of privacy protection mechanisms in CPS.The existing methods also have the following shortcomings: 1)The quantitative relationship between the degree of privacy protection and the loss of co ntrol performance is ambiguous.2)The modeling generalization ability of stealthy attacks is insufficient,which makes the designed attack model can only de al with specific attack targets.3)The decoupling capability of the designed attack detection mechanism and CPS is insufficient,which makes it seriously affect the control performance and has weak migration capability.To address the above research deficiencies,based on the existing research,this thesis transforms the above problems into the tradeoff between privacy and control performance,privacy and attack detection,and privacy and estimation performance,and designs the corresponding algorithms respectively.The main work and contributions are summarized as follows:1.For the trade-off between DP and control performance in CPS,a specific form of injecting DP into CPS is given.Then,an optimal linear quadratic Gaussian controller under DP is designed.Meanwhile,the close-form expression between privacy protection level and system control performance is obtained.In addition,a stealthy attack detection scheme fusing DP and pseudo-random number generator is designed to protect the security of CPS.2.For the trade-off between balancing DP and detection performance,a model of stealthy attack under DP is formulated.An optimized attack based on the model design is derived by solving the problem of maximizing attack expectation under DP.In contrast,an auxiliary matrix-based attack detection scheme is designed against this attack to achieve effective detection of false data injection attack without degrading the system estimation performance.On the basis of this detection mechanism,a minimization of differential privacy noise scheme is designed to optimize the estimation performance by convolving measurement noise with privacy noise,which reduces the scale of privacy noise added at each moment while guaranteeing the privacy preserving level.After that,a DP noise periodic scheduling strategy is designed to establish the correspondence between privacy noise and quadratic cost function and reduce the scale of privacy noise in one period.Therefore,the negative impact of DP on the system estimation performance is systematically reduced. |
