🤖 AI

Machine and Deep Learning for Indoor UWB Jammer Localization

This paper addresses the challenge of indoor UWB jammer localization under changing environmental conditions by introducing two novel datasets and proposing a domain-adversarial ConvNeXt autoencoder that significantly improves transferability and reduces localization error compared to non-adversarial baselines.

Original authors: Hamed Fard, Mahsa Kholghi, Benedikt Groß, Gerhard Wunder

Published 2026-07-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hamed Fard, Mahsa Kholghi, Benedikt Groß, Gerhard Wunder

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a friend in a crowded, noisy room using a special flashlight that bounces off walls to tell you exactly where they are. This is how Ultra-wideband (UWB) technology works for tracking things like your keys, your pet, or even robots in a smart building. It's incredibly precise, usually pinpointing locations within the width of a few fingers. But, like any flashlight, it can be blinded. If someone stands in the room with a giant, blinding strobe light (a "jammer"), the system goes blind, and the tracking fails. This is a big problem for security; if a bad actor can jam the signal, they can hide their location or stop a security system from working.

To fix this, scientists have started teaching computers to be detectives. Instead of just looking at the signal, they use Machine Learning (ML) and Deep Learning (DL)—basically, computer programs that learn from examples—to figure out where the blinding light is coming from, even when the signal is messy. The big question is: Can these computer detectives learn to find the jammer in one room, and then instantly do the same job in a completely different room where the furniture has been moved? Usually, computers are terrible at this; they memorize the specific room they trained in and get confused when the layout changes. This paper tackles that exact puzzle: how to build a jammer-hunting AI that doesn't get lost when the environment changes.

The Detective's Dilemma: One Room vs. The Next

The researchers set up a real-life experiment to test this. They built a "jamming lab" with four receivers (like security cameras) and a robot that carried a legitimate signal. Then, they introduced a "villain"—a device that constantly blasted interference to confuse the system. They collected data in two different scenarios. First, they mapped out 52 different spots where the jammer could hide in a room with standard furniture (the "Source" room). They trained their AI models on this data, and the results were impressive. The best model, a type of algorithm called XGBoost, could guess the jammer's location with an average error of just 20.16 cm (about 8 inches). It was like a detective who could find a hidden coin in a specific living room with incredible accuracy.

But then, the researchers played a trick. They moved the furniture—two desks, a chair, and some small objects—creating a completely new layout (the "Target" room). They tested the same AI models that had mastered the first room. The result was a disaster. The XGBoost model, which was so confident before, suddenly became clueless. Its average error skyrocketed to 207.99 cm (over 6.5 feet). It wasn't just a little off; it was completely lost. This proved that standard AI models are like students who memorize the answers to a specific test but fail immediately if the questions are rearranged. The "domain shift" (the change in the room) broke them.

The Solution: The Shape-Shifting Detective

To fix this, the authors didn't just try to train the AI on more data; they changed how the AI learned. They introduced a new method called a "Domain-Adversarial ConvNeXt Autoencoder" (or A-CNT for short). Think of this as training the detective to ignore the furniture entirely and focus only on the shape of the noise the jammer makes.

Here is how it worked:

The Autoencoder: First, they taught the AI to compress the messy signal data into a tiny, efficient summary, stripping away the "noise" of the specific room layout.
The Gradient Reversal Layer: This is the clever part. They added a special "adversarial" step. Imagine a game where the AI tries to learn the location of the jammer, but a second "critic" AI tries to guess which room the data came from. The main AI is trained to fool the critic. If the critic can't tell the difference between the "Source" room and the "Target" room, it means the main AI has learned features that are the same in both places. It forces the AI to learn the universal rules of jamming, not the specific rules of one room.

The Results: A Massive Comeback

When they tested this new A-CNT system on the messy, moved-furniture room, the results were stunning. The system didn't just recover; it thrived.

The average error dropped from the disastrous 207.99 cm down to 34.67 cm (about 13.6 inches).
This was an 83% improvement over the best previous attempt that didn't use this adversarial trick.
Most importantly, the fraction of guesses that were within a safe 30 cm (about 1 foot) jumped from a pathetic 0.03 (almost never) to 0.56 (more than half the time).

The paper also compared this new method against other standard "transfer learning" tricks (like CORAL and MMD) and found that while those helped a little, the adversarial approach was the clear winner. They even checked if the AI was actually learning the location and not just the room type. By using a simple math trick on the AI's internal "brain" (the bottleneck features), they found that the AI could still separate the jammer's location into distinct zones with 99.37% accuracy, proving it truly understood the spatial layout despite the furniture changes.

What This Means

The paper concludes that while standard machine learning is great in a static world, it fails miserably when the environment changes. However, by using this adversarial "fooling" technique, we can build localization systems that are robust enough to handle real-world chaos. The authors suggest that this approach could be the key to making UWB security systems reliable in dynamic places like smart homes or warehouses where things are constantly being moved. They also note that this is just the beginning, hinting that future work could tackle 3D spaces and even more complex layouts. The takeaway is clear: to find a jammer in a changing world, you need a detective that learns to ignore the furniture and listen to the signal.

Technical Summary: Machine and Deep Learning for Indoor UWB Jammer Localization

Problem Statement

Ultra-wideband (UWB) technology offers centimeter-scale accuracy for real-time positioning, making it vital for asset tracking and intrusion detection. However, UWB systems are vulnerable to jamming attacks, which can disrupt localization services. While machine learning (ML) and deep learning (DL) have advanced the localization of legitimate tags, the specific challenge of localizing malicious jammers within indoor environments remains largely unexplored. Furthermore, existing ML/DL frameworks often fail to generalize across varying physical environments; models trained in one room layout frequently suffer severe performance degradation when deployed in a modified layout due to domain shift. This paper addresses the need for robust, domain-invariant jammer localization that can withstand changes in indoor furniture and layout.

Methodology

Data Collection and Experimental Setup

The authors introduced two novel UWB datasets collected in an indoor laboratory (3 m × 5 m) using four UWB receivers (anchors) and a legitimate transmitter mounted on a TurtleBot robot.

Source Domain: Data collected with standard furniture at 52 predefined jammer positions.
Target Domain: Data collected after moving two desks, a chair, and other objects, at 16 new, randomly selected jammer positions.
Threat Model: The jammer uses a DW3000-series transceiver to transmit continuous UWB frames at maximum regulatory power, simulating a targeted attack on the 6.5 GHz channel.
Data Volume: The source dataset contains 461,795 samples, and the target dataset contains 28,793 samples.

Feature Engineering

Two types of features were utilized:

Diagnostic Features: Numeric readings from the DW3000 transceiver registers (e.g., PHY-header errors, Reed–Solomon failures, CRC integrity, RSSI, and Ipatov peak/power metrics).
Channel Impulse Response (CIR) Features: Raw complex CIR taps (300 taps) converted into magnitude, sine phase, and cosine phase, then truncated to the first 100 taps to reduce noise.

Baseline Models

The study established comprehensive baselines using both classical ML and DL approaches:

ML: Random Forest (RF), XGBoost (XGB), and K-Nearest Neighbors (KNN).
DL: SimpleNN (residual network), ConvMixer1D, and Transformer Tabular.
Hyperparameters were optimized using Bayesian optimization (Optuna).

Proposed Framework: Adversarial Domain Adaptation

To mitigate domain shift, the authors proposed a Domain-Adversarial ConvNeXt Autoencoder (A-CNT) framework. The architecture consists of:

Denoising ConvNeXt Autoencoder: A lightweight encoder-decoder network designed to process CIR data. The encoder compresses input via convolutional blocks with layer normalization and depthwise convolutions. Gaussian noise is injected into activations to promote robustness.
Gradient Reversal Layer (GRL): Inserted between the encoder and a domain classifier. During backpropagation, it multiplies gradients by $-\lambda$ , forcing the encoder to learn features that are indistinguishable between source and target domains.
Training Phases:
- Pre-training: Unsupervised reconstruction of source data.
- Joint Adversarial Alignment: Unsupervised training on both source and target data to minimize reconstruction loss and domain classification loss (via GRL).
- Fine-tuning: Supervised regression on labeled target data to predict coordinates, while maintaining the domain classifier as a regularizer.

Key Contributions

Novel Datasets: Introduction of two UWB datasets (source and target) capturing diagnostic features and raw CIR taps in identical rooms with different layouts.
Comprehensive Baselines: Establishment of classification and regression benchmarks using diagnostic features, supported by automated hyperparameter optimization and interpretability analysis (SHAP, mutual information, eta-squared).
Adversarial Framework: Adaptation of a domain-adversarial ConvNeXt autoencoder to align CIR-derived features across domains, specifically targeting jammer localization.
Comparative Analysis: Rigorous comparison of the proposed A-CNT against non-adversarial transfer learning (CNT) and unsupervised domain adaptation baselines (CORAL, MMD).

Experimental Results

Source Domain Performance

On the source dataset, classical ML models outperformed deep learning models in both classification and regression tasks:

Classification: Random Forest achieved the highest F1-macro score of 0.95 with a runtime of 3 minutes.
Regression: XGBoost achieved the lowest mean Euclidean error of 20.16 cm and an $F_{\le30cm}$ (fraction of predictions within 30 cm) of 0.80.

Domain Shift Impact

When source-trained models were applied to the modified target layout without adaptation, performance degraded catastrophically:

XGBoost's mean error increased tenfold from 20.16 cm to 207.99 cm.
$F_{\le30cm}$ dropped from 0.80 to 0.03.
Deep learning models (ConvMixer1D, SimpleNN) also failed, with mean errors exceeding 200 cm and negative $R^2$ values.

Domain Adaptation Performance

The proposed A-CNT framework significantly restored performance:

Mean Euclidean Error: Reduced to 34.67 cm (compared to 207.99 cm for the baseline).
Accuracy: $F_{\le30cm}$ improved to 0.56.
Comparison: This represents a 77% improvement over non-adversarial transfer learning (CNT, 148.02 cm error) and an 83% improvement over the best source-trained baseline.
Feature Alignment: t-SNE visualizations and logistic regression tests on bottleneck embeddings confirmed that the adversarial alignment preserved fine-grained spatial discriminability while aligning domain distributions.

Significance and Claims

The paper claims that adversarial feature alignment is essential for robust and transferable indoor jammer localization in dynamic environments. While traditional ML models are effective in static settings, they are highly susceptible to environmental changes. The study demonstrates that a domain-adversarial approach using a ConvNeXt autoencoder can effectively mitigate domain shift, restoring localization accuracy to near-source levels despite significant layout changes. The authors emphasize that this work fills a critical gap in the literature by focusing specifically on jammer localization rather than benign tag localization, and by addressing the practical challenge of generalization across physical environments. The code and datasets are made publicly available to facilitate further research.