This paper proposes a reinforcement learning framework for dynamic secret protection in discrete event systems. We formulate the secret protection problem as a reinforcement learning task over automata. Compared with classical supervisory control theory, the proposed framework provides an automated approach for dynamic secret protection in the systems without requiring the explicit construction of a security automaton. During training, the agent explores protection and non-protection actions at each state. State-action pairs are penalized when secret states are reached with insufficient protection, and rewarded when security requirements are satisfied, while protection costs are also taken into account. Through interaction with the environment, the agent gradually learns an optimal protection policy. We implement and compare Q-learning and SARSA algorithms. Experimental results show that both methods achieve a high success rate and produce feasible protection policy. However, SARSA demonstrates higher training efficiency and better robustness.
Dynamic Secret Protection in Discrete Event Systems via Reinforcement Learning / Ren, J., Liu, R., Mangini, A.M., Fanti, M.P.. - (2026), pp. 2122-2127. (12th International Conference on Control, Decision and Information Technologies, CoDIT 2026 Polytechnic University of Bari, Orabona Street 4, ita 2026) [10.1109/CoDIT70676.2026.11631019].
Dynamic Secret Protection in Discrete Event Systems via Reinforcement Learning
Ren J.;Liu R.;Mangini A. M.;Fanti M. P.
2026
Abstract
This paper proposes a reinforcement learning framework for dynamic secret protection in discrete event systems. We formulate the secret protection problem as a reinforcement learning task over automata. Compared with classical supervisory control theory, the proposed framework provides an automated approach for dynamic secret protection in the systems without requiring the explicit construction of a security automaton. During training, the agent explores protection and non-protection actions at each state. State-action pairs are penalized when secret states are reached with insufficient protection, and rewarded when security requirements are satisfied, while protection costs are also taken into account. Through interaction with the environment, the agent gradually learns an optimal protection policy. We implement and compare Q-learning and SARSA algorithms. Experimental results show that both methods achieve a high success rate and produce feasible protection policy. However, SARSA demonstrates higher training efficiency and better robustness.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


