Reinforcement Learning-Based Multinode Interdiction in Unknown Networks

Area of Science:

• Computer Science
• Electrical Engineering
• Network Security

Background:

• Cooperative jamming in communication networks presents challenges due to unknown topology and complex decision-making.
• The credit assignment problem is significant under full-bandit feedback, hindering effective strategy development.

Purpose of the Study:

• To develop a novel reinforcement learning (RL) algorithm for multinode cooperative jamming in networks with unknown topology.
• To address the combinatorial explosion in decision-making and the credit assignment challenge.

Main Methods:

• A multiarmed bandit (MAB) framework is utilized, featuring a collective-to-individual reward allocation model.
• A jamming overlap coefficient quantifies node interdependencies, and LASSO regression estimates individual contributions.
• A triphase learning strategy (collection-construction-exploitation) balances exploration and exploitation.

Main Results:

• The algorithm progressively identifies optimal jamming combinations without prior network knowledge.
• Theoretical analysis shows a sublinear regret bound under specific conditions.
• Simulations demonstrate superior performance over existing benchmarks in cumulative regret.

Conclusions:

• The proposed RL algorithm offers a robust and adaptable solution for cooperative jamming in unknown network topologies.
• The method effectively manages complex decision-making and reward allocation challenges.
• This approach enhances network security by enabling efficient identification of disruptive node combinations.