United States

Radar-based automatic target recognition (ATR) of UAVs is critical for security and defense. UAV recognition is commonly done using micro-Doppler (μD) signatures which are characterized by Helicopter Motor Rotation (HERM) lines that appear parallel to the main UAV return and are unique per the number of and rotation rate of propellers. This paper addresses two challenges: (1) accurate synthesis of target signatures and (2) development of classifiers robust to dynamic clutter environments, which have a different distribution than the training dataset. 

The first challenge to deep learning-based ATR is a lack of adequate amounts of quality data for training. It is often not feasible to acquire data in operational areas of interest, consequently synthetic training data generation has been a topic of intense research. However, current methods do not account for sensor-specific, target-specific, or environment-specific characteristics, which may also vary with time. Generative Adversarial Networks (GANs) are an alternative way to synthesize data from a few samples of real data, which can capture such dynamic factors. However, our prior work has shown that GANs synthesize kinematically impossible target signatures, which can degrade classification accuracy. In synthetic UAV signatures, HERM lines are often distorted, duplicated, or the trajectories are corrupted. The second challenge is that data acquired in real operational environments often has a signal-to-interference-plus-noise ratio (SINR) than the training dataset which degrades target recognition.

We propose a novel physics-based approach that addresses both UAV signature synthesis and classification in dynamic environments. First, we develop a Physics-Guided Semantic Segmentation (PhG-SS) network which uses U-Net as a backbone and is trained on low-fidelity physics-based UAV simulations. To then synthesize large datasets for training which account for the inherent kinematics of UAVs, we propose a Semantic-Segmentation GAN (SS-GAN) in which PhG-SS is added as an auxiliary discriminator branch. We show that this results in synthesis of UAV signatures with kinematically accurate HERM lines. When SS-GAN data is used in training, we found that it improved the classification accuracy of a convolutional neural network (CNN) even when there is mismatch between the SINR of training and test sets. Second, we propose a Semantic-Segmentation DNN (SS-DNN) in which PhG-SS is integrated as signal conditioning prior to convolutional layers. Our results show that the combination of training the SS-DNN using SS-GAN signatures synthesized from high SINR data results in the highest classification accuracy when testing on UAV signatures with low SINR. The implications of these results are that our proposed approach can enable target recognition in new environments from which no a priori training data is available.


NAML 2025

Physics-Guided Semantic Segmentation for UAV Radar Micro-Doppler Synthesis and Classification in New Environments

radar

deep learning

technical paper

### Welcome to the 9th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. 
 Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications
* To connect machine learning researchers with experts in military requirements
* To build and strengthen collaborations within the defense research community

#### *"What we do as a Navy is a human business. Don’t pass up an opportunity to come together as a team. Don’t pass up leadership opportunities, and don’t outsource them to a PowerPoint. If you have an opportunity to come together as a team, to increase belonging, do it."*
Admiral John M. Richardson, 31st Chief of Naval Operations

Please log in to explore this event.You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

**Get Your Tickets for NAML 2025 - Complimentary to students, government, and military**

Dive into the world of machine learning research relevant to Naval applications.

Now you can have access to all the lectures & posters and even engage with the speakers (send them emails, ask questions) on the Underline platform. 

We are offering a one-time only discounted on-demand access to this NAML 2025 content **for only $99** (complimentary to students, government, and military). 
 
**To register, please click** [**here**](https://naml25ondemand.paperform.co/)

Please register!

The 9th annual workshop on Naval Applications of Machine Learning is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. The goals of NAML are to spread awareness of current machine learning research relevant to Naval applications, to connect machine learning researchers with experts in military requirements and to build and strengthen collaborations within the defense research community.

Point measurements of optical turbulence often rely on sensitive instruments that can be challenging to deploy in various environments, and existing turbulence models may not always perform well under these conditions. In a recent experiment at the Naval Postgraduate School, optical turbulence measurements were collected using a sonic anemometer along with various atmospheric parameters (e.g., air temperature, humidity, solar flux) over a thirteen-month period. Our previous work utilized these atmospheric parameters as predictors to develop a bagged ensemble of regression trees, with optical turbulence as the response variable, and employed feature selection techniques to identify the most relevant parameter set. Building on this foundation, we conducted an in-depth analysis of how varying random partition sizes affect the model’s RMSE. Overall, the predicted values from the model demonstrated strong agreement with the observed turbulence values derived from the sonic anemometer, highlighting the model's robustness and practical applicability.

Enhancing Machine Learning Prediction Models of Optical Turbulance Through Feature Selection and Partition Analysis

Malware attribution has proven challenging, as cyber forensic expertise is in short supply and agencies struggle to scale identification from cyberattack to threat actor. We present a novel deep learning approach for automated classification of malware by attribution categories, and present SoTA results on both pairwise and one-to-many classification tasks that far exceed existing work. Our model trained solely on the MOTIF benchmark achieving 80% validation classification accuracy. We achieve this purely via a Siamese model trained on an embedding of instructions in the malware disassembly. 

An existing challenge with many malware attribution approaches is the ability to scale to classification categories and datasets beyond the limited available training data. While standard regularization techniques, such dropout, remain invaluable to more general predictions, we present two major innovations for greater generality: we model the problem as a similarity metric learning problem using a Siamese architecture to achieve a general attribution embedding space, and apply novel data augmentation to our malware samples to increase the effective diversity of our dataset. 

Our unique Siamese architecture solves the open set problem, enabling automated clustering of attribution categories that are outside of the original dataset, enabling scalability to new malware that is outside of known threat actors. We further find the embedding space produced by our model is meaningful; particularly, nearby clusters by our model in the embedding space represent closely related threat actors. Our unique data augmentation also improves model performance on the MOTIF set, and the produced embeddings for a malware sample and its augmentations suggest our model is robust to slight variations in malware features meaningless to attribution.

Siamese Networks Effectively Learn Robust Features for Malware Attribution

Integrating advanced machine learning paradigms with autonomous systems has opened new frontiers in robotics. Inspired by Hapi, the Egyptian god of coordination and harmony, this project introduces HAPI-RL. This framework merges voice-command-driven operations with reinforcement learning (RL) to enable real-time, adaptive drone swarm control. By combining natural language processing (NLP) and RL, HAPI-RL transforms voice inputs into coordinated drone actions, paving the way for innovative disaster response, logistics, and surveillance applications.
At the heart of HAPI-RL is a custom-designed RL pipeline tailored to control a drone swarm operating on the ArduPilot platform. The pipeline integrates state-of-the-art RL algorithms such as Proximal Policy Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG) to train autonomous agents capable of interpreting and executing complex voice commands. Multi-agent RL frameworks like RLlib are employed for swarm coordination, enabling drones to maintain formation, avoid collisions, and optimize task allocation dynamically. This approach ensures seamless execution of coordinated tasks while adapting to dynamic environmental changes.
Voice commands are captured and processed using open-source voice-to-text frameworks like Vosk, Kaldi, and Mozilla DeepSpeech, forming the backbone of the command recognition pipeline. Preprocessing pipelines for audio signals ensure synchronization with telemetry data, allowing accurate mapping of voice inputs to actions. To enhance reliability in high-noise environments, robust signal processing algorithms are implemented, ensuring precision in command recognition.
HAPI-RL leverages simulation environments like Gazebo and Webots to train RL models in diverse scenarios. These simulations represent the drone swarm's behavior in various operating conditions. Additionally, Software-In-The-Loop (SITL) frameworks validate behaviors before deployment in real-world settings. Upon successful training, the models are deployed on edge devices such as Raspberry Pi or Nvidia Jetson to achieve low-latency, real-time decision-making. MAVLink-based integration ensures seamless communication between the edge devices and drones.
A key feature of HAPI-RL is the implementation of abort commands as high-priority interrupts, enabling immediate override of ongoing tasks in critical scenarios. This guarantees safety and operational reliability, especially in high-stakes missions. Furthermore, extensive testing under noisy and dynamic environments validates the system's robustness, pushing the boundaries of autonomous swarm control.
HAPI-RL offers a scalable, adaptive solution for autonomous drone swarms by combining voice-to-text NLP with RL-driven coordination. Its applications range from disaster relief and surveillance to precision logistics, marking a significant advancement in autonomous systems. Through its innovative design and integration of cutting-edge technologies, HAPI-RL not only honors its namesake by fostering harmony and coordination but also sets a new benchmark for intelligent swarm control.



HAPI-RL: Hierarchical Autonomous Processing and Intelligence with Reinforcement Learning for Drone Swarms

The focus of this research is to develop and assess reinforcement learning (RL) based control schemes for control of unmanned underwater vehicles (UUVs) in multi-agent dynamic docking scenarios. UUVs are essential in military, scientific, and commercial applications. Applications include intelligence, surveillance, reconnaissance, bathymetric surveying, inspection and identification, biological sensing, and more. Modern underwater vehicle guidance and control often relies upon predefined expert-designed plans or controllers. However, RL offers a more flexible rapid learning approach. RL agents learn from interactions and experiences within their environment, whether simulated or real, enhancing decision-making and optimizing actions. This process allows them to develop behaviors and strategies given operationally complex objectives. Recent successes of reinforcement learning has seen it applied in diverse fields-from board games to aerial combat. Which demonstrates its versatility and potential for enhanced decision making. Applying these proven RL strategies could revolutionize the performance of UUVs, enabling them to navigate and operate in complex underwater environments with increased precision and maneuverability. The Proximal Policy Optimization (PPO) algorithm is employed to train RL-based controllers for 6-DOF REMUS 100 UUV simulation models in two scenarios: dock-pursuit and dock-loitering. Dock-pursuit focuses on intercepting a moving platform while avoiding collisions and minimizing control effort, whereas dock-loitering prioritizes maintaining optimal field-of-view coverage and strategic positioning relative to the docking platform. The simulation integrates a collision detection system, an alpha-beta filter for state estimation and imperfect observations, environment/state randomness, and a field-of-view sensor. Results from numerous trainings and simulations for various different scenarios is presented including high- and low-level control. Finally, the work acknowledges limitations, including the absence of sim-to-real validation, and outlines potential avenues for future research. By demonstrating RL-based UUV control, this work aims to demonstrate RL's potential in advancing performance in operationally complex underwater domains.

Multi-Agent Reinforcement Learning for Control of Unmanned Underwater Vehicles in Dynamic Docking Scenarios

In the past five years, advancements in Artificial Intelligence (AI) have significantly impacted Naval operations, enhancing real-time data processing, predictive analytics, and autonomous decision-making. The integration of AI at the edge—where data is processed locally—has become crucial for improving operational efficiency, reducing latency, and bolstering security. This is particularly vital in the context of the current geopolitical landscape, characterized by peer competition and emerging contingencies. This white paper outlines the necessity and strategic advantages of deploying and integrating AI at the edge today to address the evolving challenges of modern Naval operations.

Integrating AI at the Edge Today:  Transforming Naval Strategies and Operations

Passive underwater acoustic recording technologies are essential for identifying and localizing underwater objects. However, advanced analyses of this high-volume data type are typically performed in post-processing or aboard vessels with abundant power. With the rise of autonomous maritime vehicles and underwater robotics, passive acoustic sensing is increasingly viewed as an attractive solution for environmental and situational awareness. Utilizing this data on resource-constrained edge devices, however, requires efficient real-time processing with low power consumption. Traditional multi-target tracking methods often rely on computationally intensive techniques and still frequently require manual intervention to clean and identify tracks. To address the growing need for real-time acoustic data processing, we introduce an automated multi-target tracking system designed for small volumetric four-channel hydrophone arrays. Our approach is optimized for low latency (~1.7 ms) and low power consumption (~1.3 W) on a Raspberry Pi Zero 2W. We demonstrate the system’s performance on passive acoustic recordings of diving beaked whale groups in open ocean conditions. Our method integrates onboard neural networks with post-training optimization to enhance tracking performance while minimizing computational demands. This lightweight processing system is designed for cross-platform flexibility, making it suitable for deployment on gliders, drifters, and profilers. We detail the methodology and present preliminary results, highlighting its potential for real-time autonomous underwater applications.

Listening at the Edge: Real-time Multi-Target Acoustic Tracking for Mobile Assets

Machine learning (ML)-based sonar perception is most commonly developed through the lens of offline image processing, i.e., measured sonar time series are converted into spatial acoustic intensity maps (2D or 3D) and processed, one map at a time, using neural networks with convolutional filters. While this is sufficient for many scene classification, reconstruction, and segmentation tasks, real-time 3D autonomy tasks, such as dynamic harbor navigation, are understood to benefit from both temporal memory and multi-sensor perception, due to the inherent ambiguities in sonar, the possibility of occluded agents, the 3D nature of underwater navigation, etc.

A multi-modal, self-supervised perception algorithm was designed to provide spatio-temporal awareness information to an autonomy system. The algorithm has (a) the ability to model temporal dependencies (i.e., analyze new data in the context of its previously measured surroundings and the anticipated context in the near-future), and (b) a fused, synchronized, multi-sensor perception schema, and (c) was trained in a completely self-supervised, online manner. The initial prototype was trained on Forward-Look Sonar (FLS) using simulated data from HoloOcean. The objective is to build a rich feature space that can be used to produce both interpretable outputs and also support higher-level autonomy, such as route planning, target identification,
and so on.

Multisensor Spatiotemporal Awareness

Bottlenose dolphins often conceal behavioral signs of illness until it reaches an advanced stage. Motivated by the efficacy of vocal biomarkers in human health diagnostics, we utilized supervised machine learning methods to assess various model architectures' effectiveness in classifying dolphin health status from the acoustic features of their whistles. A gradient boosting classifier achieved an 72.3% accuracy in distinguishing between normal and abnormal health states—a significant improvement over chance (permutation test; 1000 iterations, p < 0.001). The model was trained on 30,693 whistles from 15 dolphins and the test set (15%), totalled 3612 ‘normal’ and 1775 ‘abnormal’ whistles. The classifier identified the health status of the dolphin from the whistles features with 72.3% accuracy, 73.2% recall, 56.1% precision, and an 63.5% F1 score. These findings suggest the encoding of internal health information within dolphin whistle features, with indications that the severity of illness correlates with classification accuracy, notably in its success for identifying 'critical' cases (94.2%). The successful development of this diagnostic tool holds promise for furnishing a passive, non-invasive, and cost-effective means for early disease detection in the Navy’s bottlenose dolphins.

Navy Dolphin Health Classification from Whistle Features

It has widely been observed that large language models (LLMs) have a tendency to hallucinate, which includes the generation of factually incorrect statements. This limits their use in high-consequence national security applications. While verification of outputs through external knowledge sources is often the most effective method for detecting factually incorrect content, there are many situations where an appropriate database may not exist. Previous works have demonstrated that, surprisingly, the hidden states of an LLM themselves often carry signals of hallucination that can be extracted with careful analysis. We discuss progress in this direction using both toy arithmetic datasets where granular analysis can be performed and larger, real-world datasets.

Detecting lies when you don't know the truth: Catching LLM misbehavior in the absence of external knowledge sources

We will present our cutting-edge research on securing AI models against adversarial threats, focusing on trustworthy AI in defense and operational environments. Our work explores techniques to detect, assess, and remediate AI security vulnerabilities, even in low-compute and low-connectivity environments critical to naval and military applications. We will also present our approach to evaluating AI trustworthiness, safeguarding models from adversarial attacks, and enhancing resilience in real-world deployments. Attendees will gain insights into our methodologies, along with a preview of our prototype tools designed to improve AI security in constrained operational settings. This session will highlight how our research enables more secure AI adoption in mission-critical scenarios, empowering decision-makers to deploy AI confidently in high-stakes environments.

Downloads

Next from NAML 2025

Enhancing Machine Learning Prediction Models of Optical Turbulance Through Feature Selection and Partition Analysis

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from NAML 2025

Enhancing Machine Learning Prediction Models of Optical Turbulance Through Feature Selection and Partition Analysis

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads