Abstract

In this thesis we study a detection method for the false data injection (FDI) attack class on discrete-time nonlinear multi-agent systems. The thesis considers and models three types of FDI attacks on multi-agent communication channels, including sensor channel, actuator channel, and neighbouring channel. To estimate the dynamics nonlinearity of each agent, we exploit a radial basis function neural network (RBFNN). We consider a leader-follower multi-agent system, where the communication between agents is modeled with an undirected graph. We proposed the weight tuning law for the RBFNN and introduced an NN-based distributed control law for each agent. The objective of each agent is to follow the leader and maintain the desired formation along the trajectory. We used the Lyapunov stability analysis to prove the uniform ultimate boundedness (UUB) of the formation error and neural network (NN) weights matrix and show that the multi-agent system reaches the desired w while following the leader.

Moreover, we proposed a distributed attack detection method to detect the FDI attack on each agent's sensor, actuator, and neighbouring communication channel. We designed an observer to estimate the state of each agent and used its estimation to form the residual signal for each agent. Using Lyapunov stability analysis, we show that when the system is reached its desired formation, the residual signal is UUB. We obtained abound for the residual signal and considered the bound as the attack detection threshold. We also provided the attack detectability condition for each agent.


The simulation results in MATLAB and Coppeliasim simulation environment are provided to demonstrate the performance of the detection methodology, proposed distributed control law, and neural network nonlinearity estimator, including three examples.