BALLAST: Physics-Aware Moving Target Defense for Industrial Control Systems Using a Digital Twin Safety Oracle

Abstract

Industrial Control Systems run on static network configurations, which leaves an attacker unlimited time to map the plant. Moving Target Defense answers this by shuffling addresses, but almost every published scheme never checks whether the reconfiguration will itself disturb the physical process it protects, so the defense can destabilise a control loop or trip an actuator. We present BALLAST, which places a learned Digital Twin between the security policy and the plant as a safety oracle. A physics-informed LSTM surrogate forecasts the physical effect of a candidate network change, a reinforcement learning agent trades security entropy against process stability, and a connection-draining handover lets in-flight Modbus transactions complete before the old mapping is withdrawn. An explicit shield rejects any action whose predicted trajectory would leave the safe operating envelope. On the SWaT water treatment testbed, Ballast cut the success rate of an adversary's reconnaissance sweep by 85 percent and recorded no safety violation across 50 runs, a 95 percent upper bound of 5.8 percent on the violation rate, against five violations and a tracking error 19 times larger for RL-MTD, the closest published analogue to our design. Measured against the strongest baseline on the physical metrics, a threshold-triggered scheme that itself records one violation rather than RL-MTD's five, Ballast still improves the tracking error by a factor of 2.5 and removes that remaining violation. The gain costs a small and deliberate concession in raw deterrence. Twin inference meets the 50 ms budget on a workstation core and an embedded GPU, and is projected to meet it on both edge CPUs when the three envelope stages are evaluated in a single batched call; the complete decision-to-enforcement path was measured only on the workstation.