A TRILEVEL INTERDICTION MODEL FOR MANAGING RECOURSE ACTIONS UNDER NETWORK DISINFORMATION ATTACKS

Abstract

Disinformation in infrastructure networks is spreading rapidly, with severe impacts on operational costs and system performance. This work develops a trilevel Attacker–Planner–Defender (A-P-D) network interdiction model on a capacitated transshipment network, in which an attacker falsifies reported supply and demand data within a budget, a planner solves a minimum-cost flow problem on the corrupted data, and a defender performs costly recourse actions to recover true demand. The problem is formulated as a single-level mixed-integer linear programme with complementarity constraints obtained by replacing the inner two LPs with their Karush–Kuhn–Tucker conditions and linearizing each complementarity slackness with Big-M approach. To accelerate solution, a Proximal Policy Optimization (PPO) reinforcement‑learning agent is used to generate strong attacks that feed the exact solver as a warm‑start. Computational experiments across a range of network sizes, densities, and attacker budgets show that the MPCC converges quickly on small instances but becomes computationally intractable as network size grows, eventually exceeding the allotted time limit. In contrast, PPO scales favourably with problem size and consistently produces feasible attacks within available time.