Proactive Security Strategies for Optical Networks Powered by Advanced Reinforcement Learning

Global internet traffic expanded by 17.2% in 2024, with mobile devices delivering over 40% of all data transmitted (Cloudflare report). Given that optical networks serve as the critical data conduit, their security demands have expanded accordingly. The study by Zhao and Wang introduces the Graph-Entangled Security Actor-Critic (GESAC) model, a fusion of reinforcement learning and graph neural network techniques aimed at providing adaptable protection specifically tailored for the complex, disaggregated optical network environment.

A Changing Threat Environment

The threat landscape has evolved considerably. According to Check Point’s 2025 Security Report, cyberattacks increased by 44% year-over-year. The World Economic Forum highlights cyber threats as a leading potential trigger for widespread crises, with nearly 40% of survey respondents anticipating significant disruption from cyber incidents soon. More notably, attackers have shifted tactics: exploiting vulnerabilities has overtaken phishing as the primary vector, with VPN and gateway device attacks rising eightfold to account for nearly 22% of intrusions (Verizon Data Breach Investigations Report).

Optical networks face vulnerabilities in three domains. The sophistication of attack techniques has moved beyond simple denial-of-service to include physical-layer manipulations like wavelength interception. The networks’ dynamic and multi-domain topologies complicate clear security boundaries, making intrusion tracking harder. Security further requires managing the trade-off between safeguarding the network and efficiently allocating resources such as bandwidth and routing.

Designing the GESAC Approach

The GESAC framework replaces inflexible, rule-based access controls with an intelligent system that adapts as the network changes. Its cross-layer spatiotemporal graph neural network maps evolving topologies and access patterns, recognizing causal links between optical path changes and service requests to dynamically adjust protection boundaries. Federated representation learning enables policy updates to occur locally in subdomains while still contributing to a globally consistent security scheme without centralizing sensitive data.

At its core, GESAC’s policy optimization uses the Soft Actor-Critic (SAC) algorithm. This method explores a wide range of potential strategies guided by entropy, balancing multiple objectives including network security, service quality, and resilience, ensuring effective response even under coordinated attack scenarios.

Architecture and Processing Flow

The SAC algorithm is combined with a graph neural network enhanced for spatiotemporal awareness. Two critic networks in SAC reduce bias in estimating action values, while the policy network prioritizes ongoing exploration to avoid premature convergence on suboptimal strategies. Experience replay buffers and target networks underpin stability during learning updates.

The GNN integrates a temporal attention mechanism that observes topology changes over sliding time windows. Internally, its self-attention modules detect subtle dependencies between network components, revealing covert paths that attacks might exploit. A gating mechanism filters information passing between network layers, blending physical-layer features such as wavelength allocation with logical-layer states like access controls. Community detection splits the network into security domains for focused local training; federated learning then consolidates these into a unified global policy.

Data Foundation and Model Validation

The validation dataset draws from a federation of heterogeneous sources: CAIDA’s topology data, intrusion logs from the Canadian Institute for Cybersecurity Intrusion Detection Systems (CIC-IDS), and threat metadata from ITU-T standards. CAIDA’s Multiprotocol Label Switching (MPLS) topologies were filtered to retain realistic node degrees and optical path profiles, grounding the physical-layer model.

CIC-IDS attack events, including brute force and DDoS attempts, were connected to topology nodes through hashed IP addresses and timestamps to form labeled sequences reflecting realistic network access. Additional attack types like wavelength eavesdropping and multi-domain privilege escalations were modeled according to ITU-T X.1818 guidelines and systematically injected along link edges, producing multi-stage, composite attack scenarios.

Ultimately, the data encompassed 57,000 time-stamped topology graphs, 2.8 million cross-domain authentication events, and 17 attack chain classifications spanning physical and logical layers.

Results Across Security and Resource Goals

GESAC excelled at detecting physical-layer threats, achieving F1 scores of 0.931 for wavelength eavesdropping, 0.915 for path hijacking, and 0.921 for cross-domain privilege escalation. (The F1 score synthesizes precision and recall into one measure, so high values imply both accurate detection of real attacks and minimal false alarms.) These levels represent more than sixty percentage points improvement over static methods and about 45 points above support vector machine baselines.

For application-layer attacks like DDoS and brute-force logins, false positive rates remained exceptionally low at 0.9% and 0.7%, significantly outperforming alternative approaches.

In resource allocation tests, the model reduced wavelength utilization variance by nearly 59%, signifying a more evenly distributed use of optical bandwidth, and cut latency variation by over 53%, resulting in steadier and more reliable data flow. These improvements persisted under scenarios mixing heavy traffic loads and ongoing attack attempts.

Testing in Simulated Live Environments

Mininet Optical emulations spanning 50 to 500 nodes, with real attack traces injected into traffic, confirmed the model’s operational viability. Average network delays ranged from 1.8 ms in smaller setups to about 3.4 ms in mid-sized ones, not exceeding 8.2 ms for larger topologies. Wavelength utilization stayed above 68.9%, while false positive alerts consistently remained below 2.1%. Even when introducing structural disruptions, the network maintained high performance levels with only minor increases in peak delays.

Practical Deployment Realities

GESAC’s architecture is designed with operational deployment feasibility as a foremost consideration. Its modular, distributed approach allows each network domain to process its own security policies and graph encodings autonomously, exchanging updates with peer domains to synchronize protection without centralized data bottlenecks. This matches well with current multi-controller software-defined networking (SDN) paradigms.

Implementation can leverage existing network components such as SDN controllers, edge computing servers, and virtualized routing platforms, provided these devices offer moderate computational capacity. While this enhances integration potential, deployment will likely require hardware upgrades in some edge nodes, operator training to manage the new adaptive policies, and steps to ensure seamless interoperability with established monitoring and management systems.

Paths for Further Refinement

Ongoing research aims to expand the attack scenarios represented in training datasets to improve detection of emerging threat types. Improving the system’s capability to dynamically allocate resources in response to varying network states and threat levels will heighten both security and efficiency. Progress toward quicker interpretation of multi-layer topological shifts and enhancing policy diversity will help the network maintain robust defenses even amid rapid structural changes.

Conclusion

GESAC shifts optical networking from passive defense to adaptive resilience by embedding causal analysis within multi-objective reinforcement learning strategies. Its unique combination of spatiotemporal graph modeling and entropy-guided exploration supports security policies that adjust fluidly with network evolution. Experiments demonstrate that this design sustains high accuracy in threat detection, optimizes resource use, and scales effectively for large, complex network deployments.

Citation: Zhao, Z., & Wang, Y. (2025). Soft actor-critic algorithm and improved GNN model in secure access control of disaggregated optical networks. Scientific Reports, 15, Article 29358.

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.