Soft actor-critic algorithm and improved GNN model in secure access control of disaggregated optical networks

Area of Science:

• Network Security
• Optical Networks
• Artificial Intelligence
• Reinforcement Learning

Background:

• Decomposed optical networks face challenges in coordinated defense due to dynamic topology evolution and multidimensional security threats.
• Existing security policies are often static and struggle to adapt to the complexities of modern, large-scale optical network environments.
• The need for adaptive, proactive security mechanisms is critical for ensuring the resilience and quality of service in these networks.

Purpose of the Study:

• To introduce the Graph-Entangled Security Actor-Critic (GESAC) model for enhanced security and resilience in decomposed optical networks.
• To develop a spatiotemporal modeling approach that captures causal dependencies for adaptive security boundary delineation.
• To jointly optimize network security, service quality, and system resilience through intelligent access control strategies.

Main Methods:

• Utilized a cross-layer spatiotemporal Graph Neural Network (GNN) for modeling evolving network topologies and capturing causal dependencies.
• Employed the Soft Actor-Critic (SAC) algorithm integrated with entropy-guided multi-objective reinforcement learning for policy optimization.
• Validated the model on a heterogeneous dataset including network topology data, intrusion detection logs, and threat characteristics, covering 12 attack scenarios.

Main Results:

• Achieved high F1-scores (0.915-0.931) for physical-layer attack detection with a low false positive rate (0.7%).
• Demonstrated significant improvements in resource optimization, increasing wavelength utilization variance by up to 58.9% and reducing latency standard deviation by up to 57.7%.
• Showcased strong policy robustness with over 100% increase in Pareto frontier coverage and a 65% reduction in policy entropy decay rate under topological mutations.
• Exhibited excellent scalability, achieving a single-step decision latency of 25.6µs at 100,000 nodes and reducing communication overhead.

Conclusions:

• The GESAC model effectively overcomes limitations of static security policies in dynamic, large-scale optical networks.
• Integrating causal inference with game-theoretic equilibrium shifts the security paradigm from passive defense to proactive resilience.
• Provides a foundation for next-generation architectures, enabling interpretable, highly adaptive security for multi-domain collaboration and computing-network convergence.