Learning Optimal Moving Target Defense in AC Power Systems via Residual Geometry

Model-based False Data Injection Attacks (FDIAs) pose a critical threat to AC power system state estimation by injecting carefully crafted measurements that evade traditional Bad Data Detection (BDD) mechanisms. Moving Target Defense (MTD) mitigates this threat by dynamically perturbing system parameters (typically line reactances) to invalidate the attacker’s model. However, existing MTD strategies often rely on computationally expensive optimization (e.g., mixed-integer programming, ACOPF-based cost evaluation) and lack a principled metric for quantifying detection improvements of FDIAs. This paper proposes a fully differentiable, neural optimization framework for MTD that directly maximizes the geometric residual between pre- and post-MTD measurement manifolds while minimizing system disruption via surrogate cost functions. Our formulation enables fast, scalable, and explainable MTD deployment without relying on nested optimization solvers. Experiments on the IEEE 118-bus system demonstrate that our method achieves a 94.7% FDIA detection rate with comparable cost to baselines, and offers operator interpretability by calculating marginal contributions of individual line reactance perturbations.