Machine and Deep Learning for Indoor UWB Jammer Localization

AI summary

Overview

Research area: Indoor ultra-wideband (UWB) localization security — specifically, machine learning and deep learning methods for locating a malicious jammer inside a room, including under changed room layouts.

Technical level: Advanced. The paper assumes familiarity with domain adaptation, adversarial training, gradient reversal layers, autoencoders, and UWB channel impulse response (CIR) processing.

One-sentence scope: The paper introduces two new UWB jammer datasets collected under two room configurations, benchmarks classical and deep learning models on them, and proposes a domain-adversarial ConvNeXt autoencoder (A-CNT) that restores jammer localization accuracy when the room layout changes.

What This Paper Is About

UWB systems can locate objects with centimeter-scale accuracy, but they are vulnerable to jamming attacks, and a jammed device cannot simply be trusted to report where the interference is coming from. The authors ask whether machine learning models can estimate the two-dimensional position of a malicious jammer from the signals received at four UWB anchors, and — critically — whether those models keep working when the room is rearranged. They show that source-trained models collapse under this layout change, then propose an adversarial domain-adaptation method that recovers most of the lost accuracy.

Key Contributions

1. Two new UWB datasets. The authors release datasets containing DW3000 diagnostic features and raw complex CIR taps, collected in a single room on different days: a source configuration with 52 predefined jammer locations (461,795 samples) and a target configuration with 16 new randomly selected jammer locations after moving two desks, one chair, and several smaller objects outside the test grid (28,793 samples).
2. Comprehensive ML/DL baselines. Classification and regression baselines are established with Random Forest, XGBoost, KNN, SimpleNN, ConvMixer1D, and a Transformer Tabular model, using Bayesian hyperparameter optimization with Optuna and interpretability analyses (SHAP, mutual information, eta-squared).
3. A domain-adversarial ConvNeXt autoencoder (A-CNT). The framework learns compact CIR-derived features with a denoising ConvNeXt autoencoder and uses a gradient-reversal layer to align source and target feature distributions, mitigating the measured domain shift.
4. Comparative evaluation of adaptation strategies. A-CNT is compared against its non-adversarial counterpart (CNT) and two unsupervised domain adaptation baselines, CORAL and MMD, quantifying the benefit of adversarial alignment.

Main Findings

• Classical ML wins on the source dataset. Random Forest achieved the highest classification accuracy and F1-macro of 0.95 in under 3 minutes on the 52-class task, with XGBoost close behind at 0.9381 accuracy and 0.9383 F1-macro in 3.00 minutes. The Transformer reached 0.9310 accuracy in 34 minutes, SimpleNN 0.9144 in 11 minutes, and ConvMixer1D 0.7797 in 13 minutes. KNN performed poorly at 0.5041 accuracy, which the authors attribute to 52 classes diluting neighbor density.
• XGBoost gave the best source-domain regression. It achieved the lowest mean Euclidean error of 20.16 cm and median error of 10.54 cm, with F≤30cm = 0.80 and P90 of 50.84 cm in 2 minutes. ConvMixer1D and SimpleNN both had mean errors near 28 cm, and Random Forest showed strongly anisotropic errors (rmse_X = 3.23 cm versus rmse_Y = 59.90 cm), yielding a mean error of 43.15 cm and F≤30cm = 0.48.
• Domain shift is severe. When source-trained models were applied to the modified room layout, XGBoost's mean Euclidean error rose tenfold to 207.99 cm and its F≤30cm fell from 0.80 to 0.03. ConvMixer1D and SimpleNN exceeded 200 cm mean error with negative axis-wise R² values, indicating performance below a constant-mean predictor.
• Adversarial alignment recovers localization. A-CNT reduced mean Euclidean error to 34.67 cm with R² of 0.92 (X) and 0.91 (Y), restoring F≤30cm to 0.56. The authors report this as a 77% improvement over non-adversarial transfer learning and an 83% improvement over the best baseline.
• Adversarial alignment beats classical UDA. CORAL reduced mean error to 173.90 cm and MMD to 171.59 cm, while the non-adversarial autoencoder (CNT) reached 148.02 cm. The adversarial variant was far better at 34.67 cm, suggesting adversarial learning provides gains beyond shallow feature alignment.
• Learned features preserve spatial structure. Pooled bottleneck embeddings of 3,000 hold-out samples were split into five zones by K-means on ground-truth coordinates; a logistic regression classifier on those embeddings achieved an average ROC-AUC of 0.9937 and accuracy of 0.9283 under 5-fold cross-validation.
• Core diagnostic features dominate. Ranking by SHAP, XGBoost gain, mutual information, and eta-squared placed RSL first (mean rank 1.00), followed by PHE (2.25), PREJ (3.25), and CRCB (3.50). Ipatov-derived features and RSSI ranked lowest, suggesting they could be removed in future work.

Methodology in Plain English

Setting up the problem. Four UWB receivers were placed at the corners of a 3 m × 5 m indoor test grid. A legitimate UWB transmitter was mounted on a TurtleBot robot on a 1 m extension arm, and a separate DW3000-series transceiver acted as the jammer, transmitting continuous UWB frames at maximum regulatory power. At each jammer position, the TurtleBot performed a full 360° rotation while emitting packets every 10 ms, and all four receivers recorded diagnostic readings and raw CIR taps. Each measurement was labeled with the jammer's true (x, y) coordinates. This was done at 52 positions in the original room and 16 new positions after the room was rearranged.

Baselines. For the source dataset, the authors normalized the DW3000 diagnostic features with a standard scaler fitted to the training set and trained Random Forest, XGBoost, KNN, SimpleNN, ConvMixer1D, and a Transformer Tabular model. Hyperparameters were tuned separately for classification and regression using Bayesian optimization with Optuna.

The adaptation pipeline. For the domain-adaptation path, each CIR sample's 300 taps were converted into magnitude, sine phase, and cosine phase, truncated to the first 100 taps (longer sequences degraded performance), and normalized with a scaler fitted jointly to source and target. The A-CNT model is a compact ConvNeXt autoencoder with 782,211 parameters: an encoder that downsamples from 3 to 128 channels through ConvNeXt residual blocks, and a symmetric decoder. Gaussian noise with σ = 0.6 is injected into encoder activations to prevent trivial identity mappings.

Training proceeds in three phases. First, the autoencoder is pre-trained on unlabeled source data for 30 epochs at an initial learning rate of 1e-3 with linear warmup and cosine annealing. Second, joint adversarial alignment runs for 40 epochs using unlabeled data from both domains, minimizing reconstruction loss on target data and a domain-classification loss passed through the gradient-reversal layer, with the reversal strength λ following a sigmoid schedule from 0.05 to 0.2. Domain-classifier AUC rose to 0.66 before declining as alignment progressed; early stopping triggered at epoch 18 when AUC reached 0.5145. Third, fine-tuning uses labeled target data for coordinate regression over 200 epochs, unfreezing only the final encoder stage, decoder layers, final convolution, and regression head, with the domain classifier still active as a regularizer. α is annealed from 0.5 to 0.1, λ_ft from 0.0 to 0.5, and β stays constant at 1.0.

Evaluation. Regression results are reported via mean, median, and 90th-percentile Euclidean errors in centimeters, the fraction of predictions within 30 cm, and per-axis RMSE, MAE, and R². The authors also visualize domain shift with per-tap Wasserstein distance and normalized mean differences, and use t-SNE to inspect feature alignment before and after the gradient-reversal layer.

Why This Matters

Impact on research. Prior work on UWB jammer localization focused on detecting and correcting corrupted range measurements, and prior domain-adaptation work in UWB targeted benign tags rather than malicious jammers. This paper fills both gaps by releasing the first datasets (to the authors' knowledge) aimed specifically at jammer localization across changing indoor layouts, and by showing that adversarial feature alignment is necessary where classical unsupervised domain adaptation falls far short.

Real-world applications:

• Asset tracking in restricted facilities, where a jammer could hide the movement of unauthorized items.
• Intrusion detection in smart buildings, where identifying the physical source of interference helps responders locate the attacker.
• Security monitoring of UWB-based access control or contactless systems that rely on wideband ranging.
• Diagnosing and responding to deliberate interference in industrial or logistics environments that depend on real-time positioning.

Industry relevance. The finding that classical models such as Random Forest and XGBoost match or beat deep architectures on the source task — at a fraction of the training time (3.00 versus 34 minutes for the Transformer) — is directly useful for resource-constrained deployments. At the same time, the collapse from 20.16 cm to 207.99 cm under a modest furniture rearrangement is a concrete warning that static-environment benchmarks alone do not predict field performance.

Future Directions

• Extending the approach to 3D layouts, which the authors explicitly name as future work.
• Scaling to multiple rooms rather than a single room under two configurations.
• Continuous domain adaptation, so that models keep updating as the environment evolves rather than requiring a labeled target collection.
• Simplifying models by removing the low-ranked features (Ipatov components and RSSI) identified in the importance analysis, as the authors suggest this could reduce model complexity.

Target Audience

Researchers and engineers working on UWB localization, wireless security, and jamming detection, as well as practitioners applying unsupervised domain adaptation or adversarial training to RF sensing problems. Those building indoor positioning products for asset tracking or intrusion detection will find the domain-shift results and the comparison of classical versus deep models particularly relevant. Readers without a background in domain adaptation will find the methodology sections dense but the experimental narrative accessible.

Authors’ abstract

Ultra-wideband (UWB) localization delivers centimeter-scale accuracy but is vulnerable to jamming attacks, creating security risks for asset tracking and intrusion detection in smart buildings. Although machine learning (ML) and deep learning (DL) methods have improved tag localization, localizing malicious jammers within a single room and across changing indoor layouts remains largely unexplored. Two novel UWB datasets, collected under original and modified room configurations, are introduced to establish comprehensive ML/DL baselines. Performance is rigorously evaluated using a variety of classification and regression metrics. On the source dataset with the collected UWB features, Random Forest achieves the highest F1-macro score of 0.95 and XGBoost achieves the lowest mean Euclidean error of 20.16 cm. However, deploying these source-trained models in the modified room layout led to severe performance degradation, with XGBoost's mean Euclidean error increasing tenfold to 207.99 cm, demonstrating significant domain shift. To mitigate this degradation, a domain-adversarial ConvNeXt autoencoder (A-CNT) is proposed that leverages a gradient-reversal layer to align CIR-derived features across domains. The A-CNT framework restores localization performance by reducing the mean Euclidean error to 34.67 cm. This represents a 77 percent improvement over non-adversarial transfer learning and an 83 percent improvement over the best baseline, restoring the fraction of samples within 30 cm to 0.56. Overall, the results demonstrate that adversarial feature alignment enables robust and transferable indoor jammer localization despite environmental changes. Code and dataset available at https://github.com/afbf4c8996f/Jammer-Loc