Optimal control for networks with unobservable malicious nodes

Classic network optimization theory focuses on models with stochastic dynamics and all nodes being observable controllable. However, modern systems often offer limited access suffer from adversarial attacks. In this paper, we focus networks unobservable malicious nodes, where the dynamics, such as external arrivals control actions of can be adversarial. We first extend existing by introducing a new maliciousness metric that constrains adversary, characterize stability region under dynamics. then propose an algorithm only operates accessible does not require direct observations show our is stabilizing long are within region. Finally, stabilizes even if estimates state erroneous, necessary sufficient conditions for to stabilizable when subjected estimation errors.