Active Jammer Localization via Acquisition-Aware Path Planning

Basic Information

• Paper ID: 2510.14790
• Title: Active Jammer Localization via Acquisition-Aware Path Planning
• Authors: Luis González-Gudiño¹, Mariona Jaramillo-Civill², Pau Closas², Tales Imbiriba¹
• Institutions: ¹University of Massachusetts Boston, ²Northeastern University
• Classification: cs.LG (Machine Learning)
• Publication Date: October 16, 2025
• Paper Link: https://arxiv.org/abs/2510.14790

Abstract

This paper proposes an active jammer localization framework that combines Bayesian optimization with acquisition-aware path planning. Unlike passive crowdsourcing approaches, this method adaptively guides mobile agents to collect high-utility received signal strength (RSS) measurements while accounting for urban obstacles and mobility constraints. To achieve this, the authors modify the A* algorithm by incorporating acquisition values into trajectory costs, proposing the A-UCB* algorithm for planning paths with high acquisition values. In simulations of realistic urban scenarios, the method achieves accurate localization with fewer measurements compared to uninformed baseline approaches, demonstrating consistent performance across different environments.

Research Background and Motivation

Problem Definition

Global Navigation Satellite Systems (GNSS) such as GPS and Galileo provide critical positioning, navigation, and timing (PNT) services for numerous applications. However, the strong dependence on GNSS makes these systems vulnerable to both unintentional and intentional interference. In particular, inexpensive personal privacy devices (PPDs) can emit high-power signals in the L-band, disrupting GNSS reception over distances ranging from tens of meters to several kilometers.

Significance

Detecting and localizing these interference sources is critical for resilient PNT operations. Traditional dedicated monitoring station deployments are costly, while crowdsourced data provides a cost-effective solution, particularly in densely populated or high-traffic areas.

Limitations of Existing Approaches

1. Passive Nature: Existing crowdsourcing methods rely on user-collected data, resulting in sparse samples clustered in redundant areas or failing to cover high-uncertainty regions.
2. Model Assumptions: Many methods assume known path loss propagation models, which fail in complex urban environments due to multipath, shadowing, and occlusion effects.
3. Lack of Adaptivity: Existing UAV methods either follow static plans or use only myopic heuristics, lacking global reasoning capabilities.

Research Motivation

The literature lacks adaptive strategies that guide data collection in a sample-efficient and environment-aware manner. This paper proposes an active localization framework to fill this gap.

Core Contributions

1. Novel Bayesian Optimization Framework: Proposes a Bayesian optimization framework for active jammer localization.
2. Acquisition-Aware Path Planning Strategy: A path planning method that balances mobility costs and acquisition benefits.
3. Sample-Efficient Strategy: An efficient method for accurately localizing jammers with minimal measurements.
4. A-UCB Algorithm*: A modified A* algorithm that incorporates acquisition values into trajectory costs.

Methodology Details

Task Definition

Localize a single static interference source in an urban environment using crowdsourced RSS measurements from static agents and adaptive active sensing from autonomous mobile agents. The goal is to estimate the jammer location by finding the global maximum of the interference power field:

$\hat{x}_J = \arg\max_{x \in \mathcal{X}} f_{true}(x; x_J)$

where $f_{true}(x; x_J)$ is the unknown interference power field and $x_J$ is the true jammer location.

Model Architecture

1. Sensing Model

Agents collect noisy RSS measurements at location $x$: $y_n = f_{true}(x; x_J) + \xi_n$ where $\xi_n \sim \mathcal{N}(0, \sigma^2)$ is additive measurement noise.

2. Prediction Model: Gaussian Process

A Gaussian Process (GP) is employed as a surrogate model for the interference power field: $f_{surr}(\tilde{x}) \sim \mathcal{GP}(0, k_\theta(\tilde{x}, \tilde{x}'))$

Feature vector $\tilde{x} = [p_x, p_y, z_x]$ includes normalized 2D coordinates and normalized building height.

The posterior mean and variance are: $\mu_n(\tilde{x}) = k_{\theta,n}(\tilde{x})^\top(K_n + \sigma^2_\eta I)^{-1}y_n$$\sigma^2_n(\tilde{x}) = k_\theta(\tilde{x}, \tilde{x}) - k_{\theta,n}(\tilde{x})^\top(K_n + \sigma^2_\eta I)^{-1}k_{\theta,n}(\tilde{x})$

A multi-scale additive kernel function is used: $k_\theta(\tilde{x}, \tilde{x}') = k^\ell_\theta(\tilde{x}, \tilde{x}') + k^s_\theta(\tilde{x}, \tilde{x}') + \sigma^2_\eta\delta_{\tilde{x},\tilde{x}'}$

3. Acquisition Function: Upper Confidence Bound (UCB)

$\alpha_{UCB}(x) = \mu(x) + \kappa \cdot \sigma(x)$ where $\kappa > 0$ controls the exploration-exploitation trade-off.

4. Path Planning Strategy: A-UCB*

The edge cost definition of the A* algorithm is modified as: $c_{x,x'} = (\lambda_{len} - \lambda_{info} \cdot \bar{\alpha})\|x - x'\|$ where $\bar{\alpha} = \frac{1}{2}(\alpha_{UCB}(x) + \alpha_{UCB}(x'))$ is the average acquisition value along the edge.

Technical Innovations

1. Acquisition-Aware Path Planning: First to directly integrate Bayesian optimization acquisition functions into path planning algorithms.
2. Multi-Scale Kernel Function: Designed to capture both short-range fluctuations and long-range trends in urban environments.
3. Constraint-Aware Optimization: Optimizes information collection while considering mobility constraints.
4. Environment-Adaptive: Enhances GP model environmental awareness through building height information.

Experimental Setup

Dataset

RSS fields are generated using MATLAB's deterministic 3D ray-tracing engine, considering two representative urban layouts:

1. Chicago Downtown: Dense urban core with narrow streets and tall buildings, strong multipath and shadowing effects.
2. Boston Common: Mixed environment with open park areas providing long line-of-sight corridors, surrounded by facades producing significant reflections and occlusions.

Dense uniform grid sampling with 2-meter spacing is employed.

Evaluation Metrics

1. Surrogate Model Error (SME): Distance between the true jammer location and the maximum of the GP posterior mean field.
2. Bayesian Optimization Error (BOE): Distance between the true location and the grid point with the highest RSS value actually sampled.

Comparison Methods

1. *A-UCB (δ = ∞)**: Upper bound method with unlimited path length.
2. *A-UCB (δ = 50)**: Proposed method with limited path budget.
3. Random Motion (RM): Uniform random movement along four cardinal directions.
4. Random i.i.d. Sampling (RIS): Uniform random sampling from the feasible set.

Implementation Details

• 100 independent trials
• $b_n = 2$ RSS measurements collected per BO iteration
• Initial crowdsourced samples $b_0 = 35$
• UCB exploration parameter $\kappa = 2$
• Measurement noise variance $\sigma^2 = 2.5$
• Maximum path length budget $\delta = 50$

Experimental Results

Main Results

Table 1: Localization Error (Median 25%-75%) across 100 Independent Trials

Key Findings

1. Rapid Convergence: A-UCB* converges rapidly after approximately 30 iterations, corresponding to fewer than 100 total measurements.
2. Environmental Adaptability: Boston Common's open sky areas show lower overall error levels, while the dense Chicago downtown environment presents greater challenges.
3. Robustness to Path Constraints: The method maintains most of its efficiency even under limited path length budget constraints.

Ablation Study: κ Sensitivity Analysis

By varying κ and measuring BOE at the 30th iteration, the following is observed:

• Very small κ (e.g., 0.1) is overly exploitative, producing large median errors and wide interquartile ranges.
• Very large κ (≥5) is over-exploratory, increasing errors and variability.
• Lowest and most stable errors occur at κ ∈ 1,3.
• κ = 2 is a robust choice within the low-error plateau.

Related Work

Main Research Directions

1. Physics-Based Methods: Fit power measurements to simple physical models, assuming known path loss propagation models.
2. Data-Driven Methods: Use neural networks and similar tools to directly learn complex nonlinear relationships between location and RSS from data.
3. UAV-Assisted Localization: Use unmanned aerial vehicles for jammer localization, though most methods lack adaptivity.

Advantages of This Work

Compared to existing work, this paper is the first to propose an active localization framework combining Bayesian optimization with acquisition-aware path planning, enabling adaptive data collection guidance in complex environments.

Conclusions and Discussion

Main Conclusions

1. The proposed acquisition-aware path planning strategy consistently outperforms uninformed baselines in convergence speed and localization error.
2. The framework maintains efficiency even under mobility constraints, highlighting its applicability in real-world deployments.
3. The method demonstrates robust performance across different urban environments.

Limitations

1. Single Jammer Assumption: The current framework focuses on single static jammer localization.
2. Grid Discretization: Discretization of the continuous domain may affect localization accuracy.
3. Environmental Prior Knowledge: Assumes obstacle locations are known.

Future Directions

1. Multi-Jammer Extension: Extend to multi-jammer scenarios using appropriately designed multimodal surrogate models.
2. Alternative Probabilistic Surrogates: Explore probabilistic surrogate models beyond GPs.
3. Real-World Deployment Validation: Verify method effectiveness in real urban environments.

In-Depth Evaluation

Strengths

1. Strong Methodological Innovation: First to combine Bayesian optimization with acquisition-aware path planning; novel technical approach.
2. Solid Theoretical Foundation: Based on Gaussian process uncertainty quantification and UCB acquisition functions with solid theoretical basis.
3. Comprehensive Experimental Design: Uses realistic urban scenarios, multiple baseline comparisons, and complete statistical analysis.
4. High Practical Value: Addresses the important real-world problem of GNSS interference.
5. Reasonable Algorithm Design: The A-UCB* algorithm cleverly integrates information value into path costs.

Weaknesses

1. Limited Simulation Validation: Verification only in simulation environments; lacks real-world experiments.
2. Missing Computational Complexity Analysis: No analysis of computational overhead for GP training and path planning.
3. Insufficient Parameter Sensitivity Analysis: Limited sensitivity analysis beyond κ for other hyperparameters.
4. Limited Comparison Methods: Lacks comparison with other active learning or Bayesian optimization methods.
5. Scalability Concerns: Grid discretization scalability in large-scale environments is questionable.

Impact

1. Academic Contribution: Provides new insights for combining active sensing and path planning.
2. Application Prospects: Broad application potential in UAVs, robot navigation, environmental monitoring, and other fields.
3. Reproducibility: Clear method description and detailed experimental setup facilitate reproduction.

Applicable Scenarios

1. Urban Environmental Monitoring: Suitable for signal source localization in urban environments.
2. Emergency Response: Rapid jammer localization to protect critical infrastructure.
3. Autonomous Robot Navigation: Provides environmental sensing and path planning capabilities for mobile robots.
4. UAV Applications: UAV execution of search and rescue, environmental monitoring, and other tasks.

References

The paper cites 21 relevant references covering multiple domains including GNSS systems, jammer localization, Bayesian optimization, Gaussian processes, and path planning, providing sufficient support for the theoretical foundation and technical innovations.

Overall Assessment: This is an excellent paper with strong technical innovation and high practical value. The authors cleverly combine Bayesian optimization with path planning to propose a novel method for solving the important practical problem of GNSS jammer localization. While there is room for improvement in real-world validation and computational complexity analysis, the overall quality is high and the work has significant academic and practical value for related fields.