6G Sensing Security: Distributed Game-Theoretic RL for Urban Beamforming and Attacker Detection

Technical Summary: 6G Sensing Security via Distributed Game-Theoretic RL

Problem Statement
In next-generation 6G Integrated Sensing and Communication (ISAC) systems, the dual functionality of wireless infrastructure creates new security vulnerabilities. Specifically, in urban environments, active attackers can manipulate beamforming directions by providing falsified channel-state or beam-quality feedback. This manipulation aims to increase interference and mislead transmitters (TXs) into directing their main beam lobes toward the attacker rather than legitimate users. The core challenge is to detect such active attackers and mitigate their impact on system capacity and security within a dynamic, uncertain environment characterized by complex urban propagation (Line-of-Sight and Non-Line-of-Sight conditions).

Methodology
The authors propose a hybrid framework integrating Reinforcement Learning (RL), Bayesian inference, and Game Theory to address these security challenges. The methodology operates through the following components:

1. System Model:

   • Scenario: A 3GPP-based ray-tracing simulation of an urban micro (UMi) environment in Dortmund, Germany, operating at 28 GHz (FR2 mmWave band) with 400 MHz bandwidth.
   • Hardware: Two base stations (TXs) and four receivers (RXs) equipped with Uniform Rectangular Arrays (URA).
   • Propagation: Channels are modeled using 3GPP TR 38.901 standards, incorporating path loss, shadow fading (log-normal), and small-scale fading (3GPP NR Clustered Delay Line). A Shooting and Bouncing Rays (SBR) method models multipath effects with up to one reflection and three diffractions.

2. State Representation and Belief Update:

   • The system maintains a Bayesian belief vector representing the probability ($P_j$) that each user $j$ is an attacker.
   • This belief is updated iteratively based on the deviation between predicted and observed channel capacity, utilizing an exponential likelihood derived from an anomaly score.
   • To stabilize the Q-learning process, continuous probabilities are discretized into confidence levels (5 levels per node), forming a state space defined by the suspected node ID and its confidence level.

3. Game-Theoretic Formulation:

   • Two game models are implemented to predict behavior and optimize utility: Nash and Stackelberg.
   • Utility Function: $U = \sum (1 - P_j)C_j - \beta_{risk} P_j C_j$, where $C_j$ is channel capacity and $P_j$ is the attacker probability. This formulation balances legitimate user performance against security risks.
   • Nash Model: TXs act simultaneously and independently, seeking a best-response equilibrium.
   • Stackelberg Model: A hierarchical structure where one TX acts as a "Leader" (prioritized by serving more users) and the other as a "Follower." The Leader optimizes its strategy anticipating the Follower's best response.

4. Reinforcement Learning Integration:

   • A Q-learning agent selects beamforming angles (action space: 25 discrete angles per antenna) to maximize total channel capacity and minimize interference.
   • Exploration vs. Exploitation: An $\epsilon$-greedy policy is used, but exploration is guided by the Stackelberg-Bayesian model rather than being purely random.
   • Memory Mechanisms: The study compares a memory-based approach (using Experience Replay with $(s, a, r, s')$ tuples) against a memoryless approach. The memory-based model aims to reduce temporal correlations and accelerate convergence.

Key Contributions

• Hybrid Framework: The paper introduces a novel architecture that combines Bayesian inference for uncertainty modeling, Game Theory for strategic interaction modeling, and RL for long-term optimization in ISAC security.
• Stackelberg Superiority: The work demonstrates that a hierarchical Stackelberg game structure outperforms a simultaneous Nash approach in attacker detection. The Leader-Follower dynamic reduces internal network interference, providing a clearer signal for the Bayesian algorithm to distinguish between malicious behavior and natural channel fading.
• Experience Replay in Security: The application of memory-based Q-learning (Experience Replay) is shown to significantly improve detection accuracy and system stability compared to memoryless updates in adversarial environments.
• Urban Ray-Tracing Validation: The methodology is validated using a realistic, digitized urban map (Dortmund) with ray-tracing, rather than simplified theoretical channel models.

Performance Results

• Attacker Detection:
  • The Stackelberg model achieved an average detection accuracy of 69.41%, significantly outperforming the Nash model at 38.38%.
  • In the Stackelberg scenario, the belief value for the attacker (RX3) rapidly converged to exceed 0.9, whereas the Nash model failed to distinguish the attacker with high confidence (values remained below 0.5).
• System Throughput:
  • The memory-based Q-learning approach achieved an average total channel capacity of 934 Mbps, compared to 841 Mbps for the memoryless approach (an 11.06% improvement).
  • Detection accuracy improved from 70.13% (memoryless) to 75.46% (memory-based).
• Robustness: Under the proposed framework, the system successfully identified attackers with probabilities consistently above 0.9, even in the presence of severe shadowing and multipath effects. The system recovered quickly from temporary belief drops caused by channel dynamics, demonstrating resilience.
• Capacity Recovery: Despite the presence of an attacker, the optimized beamforming allowed the total system capacity to reach approximately 2500 Mbps in later iterations, meeting 5G New Radio requirements.

Significance and Claims
The paper claims that the proposed framework effectively addresses security challenges in dynamic 6G ISAC systems by transforming the system from a simple numerical optimizer into a "risk-aware, uncertainty-driven, and behavior-sensitive decision-maker." The integration of Bayesian inference into game-theoretic utility functions allows for more accurate attacker detection compared to standard Nash approaches. Furthermore, the study establishes that incorporating memory (Experience Replay) into the RL process is critical for stabilizing learning and improving throughput in noisy, adversarial urban environments. The work provides a foundational basis for extending these methods to mobile and real-time ISAC scenarios, ensuring secure and efficient resource allocation in future networks.