⚡ electrical engineering

Active Localization of Close-range Adversarial Acoustic Sources for Underwater Data Center Surveillance

This paper proposes a real-time surveillance framework using a heterogeneous hydrophone setup and a Locus-Conditioned Maximum A-Posteriori (LC-MAP) scheme integrated with an unscented Kalman filter to achieve sub-meter localization and tracking of close-range adversarial acoustic sources targeting underwater data centers, effectively overcoming challenges like phase ambiguity and mobile receiver uncertainty.

Original authors: Adnan Abdullah, David Blow, Sara Rampazzi, Md Jahidul Islam

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adnan Abdullah, David Blow, Sara Rampazzi, Md Jahidul Islam

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 a massive underwater city made of glass bubbles, housing the world's most important data. These "Underwater Data Centers" are great because the ocean keeps them cool and safe from land-based thieves. But there's a new kind of thief: a sound burglar.

This burglar doesn't use a crowbar; they use a loudspeaker. They blast specific, high-pitched tones that make the hard drives inside the servers vibrate so violently they stop working or break. The problem? The ocean is dark and murky, so cameras can't see the burglar. And because sound travels so well underwater, the burglar can be just a few meters away and still cause chaos.

This paper presents a new "sonar detective" system to find and track this sound burglar in real-time. Here is how it works, broken down into simple concepts:

1. The Detective Team: One Stationary, One Moving

Most security systems use a huge grid of microphones (hydrophones) everywhere, which is expensive and hard to set up. This team uses a smarter, lighter approach:

The Anchor: One microphone is glued to the data center pod itself.
The Rover: One microphone is strapped to a swimming robot (an underwater drone).

Think of it like trying to find a hidden speaker in a dark room. If you stand still with one ear, you can guess the direction, but you don't know exactly how far away it is. But if you have a friend walk around the room while listening, their changing position helps you triangulate the exact spot. The robot swims around the pod, gathering clues from different angles.

2. The Clues: Time and Speed

The system listens for two specific clues to figure out where the sound is coming from:

Time Difference (TDOA): The sound hits the robot's microphone a split-second before (or after) it hits the pod's microphone. This tells the system how far away the sound is.
Speed Difference (FDOA): As the robot swims, the sound waves get slightly "squished" or "stretched" (like the change in pitch of a passing siren). This tells the system how fast the sound source is moving relative to the robot.

3. The Problem: The "Echo Chamber" Effect

Underwater is tricky. Sound doesn't just travel in a straight line; it bounces off the water's surface and the ocean floor.

The Metaphor: Imagine shouting in a canyon. You hear your voice, but you also hear it bouncing off the walls. If you try to guess where the echo came from, you might get confused.
The Paper's Fix: The system has a special "echo filter." It mathematically predicts where the sound should bounce and subtracts that confusion from the data, so the detective only listens to the direct path.

4. The "Smart Start" (LC-MAP)

Usually, when a computer tries to find something, it starts with a wild guess. If the guess is wrong, the computer gets stuck or takes forever to find the target.

The Metaphor: Imagine trying to find a lost key in a dark house. If you start guessing randomly, you might search the wrong room for hours.
The Paper's Fix: The authors created a "Smart Start" method called LC-MAP. Instead of guessing randomly, it uses the very first few sound clues to draw a "map" of where the burglar could be. It picks the most logical spot on that map to start the search. This makes the system find the target twice as fast as older methods.

5. The Results: Finding the Thief

The team tested this system in three ways:

Computer Simulations: They ran thousands of virtual scenarios. The new system found the "burglar" over 90% of the time and did it much faster than the old ways.
Robot Simulations: They tested it in a virtual underwater world with robots. The system tracked the target with sub-meter accuracy (less than 3 feet off).
Real Lake Tests: They took the system to a real lake. They used two boats: one carrying the "burglar" (a speaker) and one carrying the "detective" (the robot with a microphone). Even with the noise of boat motors and water waves, the system successfully tracked the moving speaker, proving it works in the real world.

Why This Matters

The paper claims that if a sound attack starts, the system can locate the source quickly enough to stop the attack before the hard drives are permanently damaged. It turns a passive listening post into an active, intelligent surveillance system that can spot a threat in the dark, murky depths without needing a massive army of sensors.

In short: They built a smart, two-person detective team (one fixed, one swimming) that uses sound echoes and a "smart guess" algorithm to find underwater sound attackers quickly and accurately, even when the water is noisy and full of confusing echoes.

Technical Summary: Active Localization of Close-range Adversarial Acoustic Sources for Underwater Data Center Surveillance

1. Problem Statement

Underwater Data Centers (UDCs) offer advantages in cooling and physical security but introduce unique vulnerabilities to acoustic injection attacks. Unlike terrestrial facilities, the nitrogen-filled pressure vessels and surrounding water of UDCs conduct sound efficiently, allowing narrowband acoustic signals to induce mechanical vibrations in Hard Disk Drives (HDDs). These vibrations can cause I/O errors, throughput degradation, and permanent system crashes within approximately 85 seconds.

While prior work has focused on detecting these attacks via internal Position Error Signals (PES), it lacks the capability to localize and track the adversarial source in 3D space. Conventional Sound Source Localization (SSL) methods are ill-suited for this specific scenario due to several constraints:

Non-cooperative, Narrowband Sources: Adversarial signals are tonal and lack synchronization, causing phase ambiguity in Time-Difference-of-Arrival (TDOA) estimation.
Scalability Issues: Deploying large, synchronized hydrophone arrays across multi-pod offshore facilities is logistically prohibitive.
Environmental Complexity: Shallow-water deployments suffer from strong surface and bed reflections (multipath) that bias measurements.
Mobile Receiver Uncertainty: Using a mobile robot for surveillance introduces uncertainty in the receiver's own state (position and velocity), which standard filters often fail to account for.
Initialization Sensitivity: Standard recursive filters (e.g., Kalman filters) often diverge or converge slowly when initialized without prior knowledge of the source state in non-linear, phase-ambiguous conditions.

2. Methodology

The authors propose an Adversarial Acoustic Source Localization (AASL) framework utilizing a heterogeneous receiver configuration: one static hydrophone mounted on the UDC pod and one mobile hydrophone carried by an autonomous surveillance robot (ASV/AUV). The system estimates the 3D position ( $p_s$ ) and velocity ( $v_s$ ) of the source by fusing TDOA and Frequency-Difference-of-Arrival (FDOA) measurements within an Unscented Kalman Filter (UKF) pipeline.

Key Technical Components:

Measurement Model:
- TDOA: Estimated via short-window cross-correlation. To address phase wrapping in narrowband signals, a temporal unwrapping algorithm is applied to maintain continuity.
- FDOA: Derived from the phase slope of the cross-spectral density between the two channels.
- Multipath Correction: A first-order image-source model is used to correct range differences for surface and bed reflections, modeling the environment as having a surface image and a bed image.
- Effective Covariance: The measurement covariance is inflated to account for the uncertainty in the mobile receiver's state (GPS drift, navigation errors), preventing the filter from becoming overconfident.
LC-MAP Initialization Scheme:
- A novel Locus-Conditioned Maximum A-Posteriori (LC-MAP) scheme is introduced to generate an informed prior for the UKF.
- Instead of random or geometric midpoint initialization, LC-MAP samples candidate positions along the hyperbolic locus defined by the initial TDOA measurement.
- It evaluates the likelihood of these candidates under the joint TDOA-FDOA model, weighted by a Fisher Information Matrix (FIM) term to favor geometrically well-conditioned regions. This ensures the filter starts with a physically plausible state, accelerating convergence.
State Estimation:
- The UKF propagates the source state using motion models (Constant Velocity, Constant Acceleration, Constant Turn Rate).
- The filter recursively fuses the multipath-corrected TDOA and FDOA measurements to refine the source trajectory.

3. Key Contributions

The paper outlines four primary contributions:

Scalable Heterogeneous Architecture: A two-hydrophone system (one static, one mobile) that provides wide-area coverage and 3D tracking without the logistical overhead of dense, synchronized fixed arrays.
LC-MAP Initialization: A novel scheme that fuses nonlinear TDOA-FDOA observations with geometric constraints to generate physically consistent priors, significantly improving convergence speed and success rates compared to standard initialization methods.
Robust Filtering Formulation: A joint TDOA-FDOA filtering approach that explicitly accounts for shallow-water multipath bias and mobile receiver state uncertainty, enhancing robustness in realistic deployment scenarios.
Comprehensive Validation: The framework is validated through Monte Carlo simulations, Gazebo-ROS physics simulations, and real-world open-water field trials, demonstrating performance across diverse attack scenarios (static, moving, turning).

4. Experimental Results

The evaluation utilized Monte Carlo simulations, Gazebo-ROS physics simulations, and 15 field trials in a shallow-water lake environment.

Initialization Performance: In Monte Carlo simulations across four motion models (static, constant velocity, constant acceleration, constant turn rate), the LC-MAP scheme achieved >90% success rates in most scenarios. It reduced the average convergence time by nearly 50% compared to a naive midpoint initialization and outperformed Random Sphere and TDOA-LS baselines.
Localization Accuracy:
- Gazebo Simulations: Achieved sub-meter Root Mean Square Error (RMSE) for both static and mobile sources.
- Field Trials: In real-world shallow-water tests, the compensated filter (incorporating multipath and receiver uncertainty) reduced the RMSE from 1.89 m to 1.37 m compared to the uncompensated nominal filter.
Scalability Comparison: Compared to fixed-array baselines (OSLS and SDP) for an 8-pod facility, the proposed system reduced the required number of receivers from 48–64 fixed sensors to just 9 sensors (8 static + 1 shared mobile), while simultaneously estimating velocity (which fixed arrays typically cannot do).
Ablation Studies:
- Multipath: Ignoring multipath reflections resulted in a 9% convergence rate and large errors (avg RMSE ~25.7 m). The multipath-aware model achieved 100% convergence with an avg RMSE of 0.98 m in simulation.
- Receiver Uncertainty: Filters that compensated for mobile receiver state uncertainty remained stable (RMSE < 3 m) even with position errors up to 70 cm, whereas uncompensated filters degraded rapidly.

5. Significance and Claims

The paper claims to establish a geometry-aware, real-time approach for acoustic threat localization that advances the autonomous surveillance capabilities of underwater infrastructure.

Practical Feasibility: By demonstrating successful localization in real-world field trials with noisy measurements, the work proves that active surveillance of UDCs is feasible without requiring impractical sensor densities.
Defensive Utility: The framework enables rapid activation of defensive countermeasures (e.g., traffic isolation, threat neutralization) within the critical 85-second window before permanent hardware damage occurs.
Generalizability: While focused on UDCs, the authors state the framework provides a scalable foundation for acoustic surveillance of other offshore assets, such as underwater sensor networks and communication relays.

The authors remain modest regarding limitations, acknowledging that the current multipath model is first-order and that real-world performance is bounded by geometric diversity, receiver navigation accuracy, and environmental interference. They identify future work in deep-water deployment, observability-driven path planning, and matched-field processing for higher-order multipath mitigation.