Many security and privacy requirements in cyber-physical systems are inherently relational, involving comparisons across multiple system trajectories, rather than properties of single trajectories. These requirements, known as hyperproperties, naturally arise in information-flow security, and include opacity, noninterference, and K-anonymity. Since standard temporal logics are limited to single-trajectory specifications, HyperLTL has emerged as a powerful tool for specifying such relational requirements. This paper presents a control synthesis framework for enforcing a class of HyperLTL security specifications in continuous-time control-affine systems using control barrier functions (CBFs). The proposed approach constructs CBFs over an augmented state space that captures the multiple trajectories quantified in the HyperLTL formula. The resulting barrier conditions are imposed as constraints in a quadratic program, enabling real-time controller synthesis that enforces hyperproperties. The framework provides a unified approach for integrating information-flow security requirements into continuous-time controller synthesis.We show how representative security notions can be encoded and enforced within this framework, and demonstrate its effectiveness through a robotic case study.
This paper develops a regret-based analysis of continuous-time optimization in closed-loop control systems, connecting online learning objectives with control-theoretic requirements. Two constrained update mechanisms are studied: a projected dynamical system (PDS), which projects the instantaneous descent velocity onto the tangent cone of the feasible set, and a projected gradient flow (PGF), which relaxes a finite projected-gradient step into continuous time. For convex online losses, the analysis establishes distinct regret upper bounds and clarifies the limiting relation between these mechanisms. The framework is applied to closed-loop residual model learning, and numerical examples illustrate key properties of both approaches.
In this paper, we address the problem of nonlinear finite-dimensional observer-based boundary control design for reaction-diffusion PDEs. By employing a modal decomposition approach, we design a nonlinear output-feedback law based on a finite-dimensional linear observer. The well-posedness of the closed-loop system is analyzed using a control approximation, and its global asymptotic stability is established by constructing an appropriate Lyapunov function. The proposed finite-dimensional observer-based controller exhibits superior performance compared to a linear controller in closed-loop settings, particularly in mitigating the spillover effect. Numerical simulations are presented to support the theoretical results.
Attack detection in Cyber-Physical Systems (CPS) is hindered by adversaries capable of preserving the local plausibility of measurements while exploiting dependencies among coupled subsystems. This study proposes the Knowledge-guided Heterogeneous Graph Network (KHGN), a neuro-symbolic architecture that represents the CPS as a typed heterogeneous graph, integrating three complementary dimensions: temporal sensor dynamics, physical topology of the system, and domain knowledge formalized through an ontology. The effectiveness of the approach is evaluated on the Secure Water Treatment (SWaT) industrial testbed for multi-stage attack detection. Across five independent runs, KHGN achieved a macro F1 of 0.928 ± 0.006 and exceeded the HGT-only variant by 14.3 percentage points. Limited-label and leave-one-attack-out evaluations further assessed performance under scarce supervision and for attacks excluded from training.
This paper investigates the event-triggered self-organized fencing control problem for second-order multi-agent systems subject to external disturbances and an unknown time-varying target. Unlike prescribed fencing configurations, a self-organized formation is generated through potential-field regulation rather than prescribed relative positions. Based on the velocity compensation function, a regulation potential function is constructed to drive the agents towards the target while the collision and obstacle avoidance potentials are further incorporated to guarantee the local safety. A dynamic controller is then developed where a smooth reinforcement learning echo state network (RL-ESN) compensation function is introduced to attenuate bounded disturbances. Furthermore, an event-driven mechanism with a safety-regulation function is designed to guarantee the system safety and reduce the unnecessary data transmission between the controller and actuators. The rigorous theoretical analysis illustrates that the regulation and tracking error signals are uniformly ultimately bounded. The simulation results demonstrate the effectiveness of the proposed method.
