PaperlayerView exact revision on arXiv

Source-linked AI summary

On Distributed Sensing and Adversarial Localization of Narrowband Acoustic Sources for Offshore Data Center Monitoring

Adnan Abdullah, David Blow, Harsha Gundala Sureshbabu, Dongyoon Joung, Sara Rampazzi, Md Jahidul Islam

arXiv:2610.34393v1cs.RO

TL;DR

Sparse receiver networks face ambiguous TDOA measurements when adversarial sources emit narrowband tones, complicating acoustic localization around data centers. The report uses mobile anchor constraints to initialize particle-filter tracking with quality-adaptive measurements. It reports 91% initial localization success in simulation and tracking error below 2 m in 9 of 10 field trials.

  • Problem

    Narrowband emissions and sparse sensing create phase-aliased, ambiguous TDOA localization conditions for acoustic surveillance of distributed infrastructure.

    Source: §1 · p. 2

  • Method

    A mobile receiver forms short-baseline anchor pairs for initial localization, then a particle filter tracks source position and velocity using a live anchor and quality-weighted fixed-pair measurements.

    Source: §1 · p. 3

  • Results

    91% initial localization success in Monte Carlo simulations, while 9 of 10 open-water trials maintained tracking error below 2 m.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The investigations characterize the potential of distributed acoustic surveillance and the practical importance of receiver geometry and measurement reliability in offshore deployments.

    Source: §1 · p. 3

    Independent receiver clocks, wireless data loss, and multipath introduce sensing errors; coarse alignment and initial calibration do not guarantee long-term phase coherence.

    Source: §4.6 · p. 15, §4.6.1 · p. 16

Abstract

from arXiv · show

This report investigates an acoustic surveillance framework for localizing and tracking narrowband adversarial acoustic sources that threaten offshore data centers. Unlike conventional sound source localization (SSL) methods that rely on broadband signals and compact receiver arrays, the considered scenario involves narrowband emissions and sparse sensors. In such scenarios, the observed time-difference-of-arrival (TDOA) measurements are aliased and produce an ambiguous solution space, rendering conventional localization methods ineffective. To address this challenge, we integrate a mobile receiver that temporarily forms anchor pairs with selected fixed receivers, constraining the feasible solution space. The proposed two-stage pipeline first estimates the initial source position using anchor constraints, then initializes a particle filter that jointly tracks source position and velocity using one live anchor measurement and quality-adaptive fixed-pair measurements. We evaluate the proposed method through Monte Carlo acoustic simulations, ROS-Gazebo robotic demonstrations, and open-water field experiments. Our framework achieves a localization success rate of 91% in simulation, while 9 of 10 field trials show tracking error below 2m under receiver-position uncertainty and noisy sensing conditions. These findings demonstrate the feasibility of passive acoustic surveillance for protecting large-scale infrastructure against acoustic threats.

1 Introduction

Narrowband acoustic attacks can threaten distributed computing infrastructure, while sparse receivers and phase aliasing make source localization challenging. The report introduces mobile anchor constraints and a two-stage localization and tracking framework, evaluated in simulation, robotic demonstrations, and open water.

1 Introduction

  • Acoustic attacks can induce mechanical vibrations in sensitive hardware such as HDDs, with prior studies reporting I/O disruption and data unavailability.This motivates locating the acoustic source to support monitoring and protection of distributed computing facilities.Source: §1 · p. 2
  • Narrowband tonal emissions can make TDOA measurements phase-aliased, leaving multiple feasible source locations; sparse sensing is also needed because compact arrays are costly to scale and centralized arrays face attenuation and multipath.The report frames the problem around sparse distributed sensing, phase aliasing, and receiver-dependent channel degradation.Source: §1 · p. 2
  • A mobile receiver forms short-baseline anchor pairs to obtain unaliased geometric constraints, then anchor-constrained initialization and alias-aware refinement initialize particle-filter tracking with quality-adaptive measurement weighting.The filter jointly tracks source position and velocity; the initialization uses two independent anchor constraints while the source is assumed approximately stationary.Source: §1 · p. 3
  • 91% initial localization success in Monte Carlo simulations, and 9 of 10 open-water trials maintained tracking error below 2 m.The results also report that reliability depends strongly on anchor measurement accuracy and that quality-adaptive weighting improves robustness under receiver-channel degradation.Source: §1 · p. 3

2 Background & Preliminaries

The section establishes the source and receiver model, explains how narrowband phase aliasing creates ambiguous TDOA localization, and motivates mobile sensing as a way to constrain solutions in sparse networks.

2 Background & Preliminaries

  • The surveillance model combines known-position distributed receivers with a movable receiver and estimates an adversarial source’s position and velocity in two or three dimensions.The mobile receiver approaches static receivers to form temporary anchor pairs.Source: §2 · p. 3

2.1 Acoustic Injection Attacks in Data Centers

  • A modeled data-center adversary emits a narrowband tone near a structural resonance, which can induce vibrations that disturb HDD operation and lead to system failures.The passage notes stronger acoustic coupling in offshore and underwater data centers.Source: §2.1 · p. 3

2.2 TDOA Measurement and Phase Aliasing

  • Narrowband range differences are observed modulo the wavelength, producing feasible alias branches that leave a continuous set of source locations and make a single pair insufficient for unique localization.The number of branches grows approximately with 2b_ij/λ, so wider receiver spacing or higher frequency increases spatial aliasing.Source: §2.2 · pp. 4–5

2.3 Prior Works & Motivation for Mobile Sensing

  • Prior large-area approaches often assume compact arrays and sufficient signal bandwidth, while sparse single receivers scale more readily but suffer severe tonal-signal aliasing and variable signal quality.The proposed mobile sensor uses high-SNR static receivers as anchors, then incorporates other pairs through alias-aware residuals and channel-quality weights.Source: §2.3 · p. 5

3 Localization and Tracking Method

The framework combines quality-adaptive weighting with mobile anchor constraints to resolve narrowband localization ambiguity, then tracks the source with a particle filter. Stage 1 traces and refines candidate positions on the anchor-constrained curve; Stage 2 estimates position and velocity from live and distributed measurements.

3.1 Quality-Adaptive Pair Weighting

  • Receiver quality is mapped from narrowband SNR to [qmin, 1], and each fixed-pair reliability takes the lower quality of its two channels.Weights are recomputed per audio frame; anchor measurements retain unit weight.Source: §3.1 · p. 6

3.2 Stage 1: Initial Localization

  • Stage 1 assumes the source is stationary while the mobile receiver sequentially visits selected static receivers to acquire anchor measurements.The stage proceeds through anchor constraints, curve tracing and sampling, then candidate scoring and refinement.Source: §3.2 · p. 6

3.2.1 Anchor Constrained Solution Space

  • Two independent anchor constraints reduce the initial candidate set to a one-dimensional curve, which aliased fixed-pair measurements then help resolve.In 3D, two independent constraints intersect along a curve; the anchor constraints alone do not determine the source’s position along it.Source: §3.2.1 · p. 7

3.2.2 Intersection-Curve Tracing

  • Predictor-corrector tracing samples the anchor-consistent curve in both directions until the trace leaves the monitored region, projection fails, or the step limit is reached.The resulting samples reduce the search from the full-dimensional region to a one-dimensional candidate set.Source: §3.2.2 · pp. 7–8

3.2.3 Candidate Scoring and Alias-Aware Refinement

  • A bounded Gaussian score ranks curve samples using reliable fixed-pair measurements, and the top K candidates are refined with anchor and fixed-pair residuals.Keeping multiple candidates accommodates local maxima from wrapped residuals; refinement allows small deviations from the anchor curve to absorb anchor-measurement noise.Source: §3.2.3 · pp. 8–9

3.3 Stage 2: Temporal Tracking

  • Stage 2 initializes particles around the Stage 1 estimate with near-zero velocities and uses a constant-velocity transition model to track position and velocity.The measurement likelihood combines one live unaliased anchor measurement with wrapped fixed-pair measurements; quality weights temper low-SNR pairs, while the anchor term and motion prior keep particles on the Stage 1 branch.Source: §3.3 · p. 9
  • Two independent anchors reduce the initial search to a one-dimensional intersection curve, avoiding the costly global search left by a single anchor or many aliased modes.The design rationale contrasts this approach with dense grid search, globally initialized particle filtering, and graph-based phase-consistency methods.Source: §3.3 · p. 10

4 Evaluation

The evaluation uses Monte Carlo acoustic simulations, ROS-Gazebo robotic simulations, and open-water field experiments. It reports initial localization and tracking metrics, with tracking error conditioned on successful initialization.

4 Evaluation

  • The evaluation spans Monte Carlo acoustic simulations, ROS-Gazebo robotic simulations, and open-water field experiments.The study examines randomized receiver geometries, noise, source motion, robotic surveillance scenarios, and lake trials.Source: §4 · pp. 10–11
  • Initial localization error is measured after Stage 1, while tracking error is time-varying and reported only for trials with successful initialization.Success rate is the percentage of trials with initial localization error below 1 m.Source: §4 · p. 11
  • Table 2 compares localization and tracking performance by number of anchor constraints; tracking error is reported only for trials with initial localization error below 1 m.Initial localization error is summarized over all trials as mean±std.Source: Table 2 · p. 11

4.1 Monte-Carlo Simulation

Monte Carlo trials test localization and tracking across randomized receiver geometries and source motion. Two independent anchors substantially improve initialization reliability under narrowband phase ambiguity.

4.1 Monte-Carlo Simulation

  • < 5 cm steady-state tracking error with limited drift was observed in representative trials across randomized receiver placements and source trajectories.Trials randomized static-receiver locations, selected anchors, initial source position, and motion direction.Source: §4.1 · p. 11
  • 91% success rate and 0.72 m mean ILE were achieved with two anchors, versus 12% and 4% success rates with one and no anchors.Two anchors reduce the candidate set to a 1D intersection curve; one anchor leaves a 2D surface.Source: §4.1 · p. 11
  • Across all 36 ordered anchor pairs, tracking remained close to ground truth after successful initialization, though initialization quality varied with geometry and measurement quality.Source initialization and motion direction were randomized.Source: §4.1 · p. 11

4.2 Robotic Simulation in Gazebo-ROS

ROS-Gazebo demonstrations evaluate tracking across floating and submerged infrastructure configurations. Tracking RMSE stays below 1 m across six trajectories, while wide paths expose receiver-coverage and geometry weaknesses.

4.2 Robotic Simulation in Gazebo-ROS

  • The demonstrations cover a rail-mounted receiver at a single barge, an eight-barge floating facility, and an eight-pod submerged facility.The distributed floating and submerged cases use USVs and UUVs, respectively, with local and facility-wide source trajectories.Source: §4.2 · p. 11, Figure 5 · p. 13
  • Figure 4 tests all 36 anchor-pair choices with randomized static-receiver locations, source initialization, and motion direction.The setup is intended to demonstrate geometric robustness.Source: Figure 4 · p. 12
  • In the multi-barge scenario, a USV traveling at a nominal 1 m/s reaches two consecutive anchor locations within 10 s.The reported interval is compatible with the considered threat model, where the source remains near one target long enough to maximize damage.Source: §4.2 · p. 13
  • Tracking RMSEs remained below 1 m in all six evaluated trajectories.Localized attacks had lower tracking error; wide trajectories showed larger deviations in regions with weaker receiver coverage or less informative geometry.Source: §4.2 · p. 13
  • Receiver placement should be designed jointly with the expected patrol region to maintain signal strength and geometric observability across the facility.Localized attacks had lower tracking error because nearby receivers provided stronger signals and better geometry.Source: §4.2 · p. 13

4.3 Open-Water Field Validation

A custom wireless receiver network and mobile platform were tested in a shallow-water lake. The framework successfully located and tracked the source in 9 of 10 trials, with errors below 2 m.

4.3.1 Receiver Hardware and Data Acquisition

  • The lake prototype used six distributed receiver modules, each digitizing microphone audio at 62 kHz and transmitting it over WiFi to a central computer.Each module combined an electret microphone, two-stage amplifier, ADC, and ESP32 board.Source: §4.3.1 · p. 13

4.3.2 Field Setup and Results

  • Figure 6 compares the estimated trajectory with a USV GPS ground-truth path for proximal and standoff attacks around a scaled floating data center.The setup includes distributed static microphones, a mobile microphone on a four-wheeled vehicle, and a USV carrying the source.Source: Figure 6 · p. 14
  • A mobile receiver on a remote-controlled vehicle moved along guard rails to approach selected static receivers and form anchor pairs.The source was a speaker mounted on a USV, with motion, emission frequency, and amplitude controlled remotely.Source: §4.3.2 · p. 14
  • In 9 of 10 trials, the method located the source and maintained tracking error below 2 m throughout the tests.The trials included proximal attacks within 5 m and standoff attacks from distances up to 15 m; standoff errors were typically higher.Source: §4.3.2 · p. 14

4.4 Sensitivity to Measurement Uncertainty

Sensitivity tests vary anchor and fixed-pair measurement uncertainty across randomized geometries. Initialization is more sensitive to anchor-measurement uncertainty, which defines the unaliased candidate manifold.

4.4 Sensitivity to Measurement Uncertainty

  • The tests model mobile-receiver position error and anchor-measurement noise together as equivalent anchor range-difference uncertainty, σa.Performance is reported using success rate and initial localization error across randomized receiver geometries and source locations.Source: §4.4 · p. 14
  • At low anchor-pair uncertainty (σa = 0.01 cm), success rate remains high at 86-92% across all σρ values, with corresponding ILE over successful trials near 10-14 cm.Source: §4.4 · p. 15
  • Reliable localization was observed when σa ≲ 0.05 cm, while increasing σρ at fixed σa produces only gradual changes in success rate and ILE.Increasing σa degrades initialization rapidly.Source: §4.4 · p. 15
  • Anchor measurements define the unaliased candidate manifold, whereas fixed-pair measurements mainly disambiguate and score points along that manifold.A poorly conditioned manifold from high σa cannot be reliably traced regardless of fixed-pair measurement quality.Source: §4.4 · p. 15

4.5 Ablation: Adaptive Receiver Weighting

The ablation tests quality-adaptive weighting when one non-anchor receiver is degraded, comparing it with uniform weighting. Adaptive weighting helps most under severe degradation and converges with uniform weighting for clean measurements.

4.5 Ablation: Adaptive Receiver Weighting

  • The test corrupts one randomly selected non-anchor receiver, whose noise affects every TDOA pair containing that channel.Uniform weighting sets every fixed-receiver pair weight to 1; adaptive weighting estimates channel quality and propagates it to pair weights.Source: §4.5 · p. 15
  • At −30 dB, adaptive weighting increases SR by approximately 10% and reduces mean tracking error from 1.09 m to 0.60 m.The benefit decreases as corrupted-channel SNR improves.Source: §4.5 · p. 15, Figure 9 · p. 15
  • Under clean measurements, both approaches converge to 92% SR and 0.03 m TE.The comparison uses identical receiver geometries and source trajectories.Source: §4.5 · p. 15

4.6 Practical Sensing Challenges

Practical phase-based TDOA measurements are affected by independent receiver clocks, wireless data loss, and multipath, which undermine temporal consistency and phase reliability. The section identifies synchronization, acquisition, and multipath handling as priorities for deployment.

4.6.1 Synchronization and Data Loss

  • Independent sampling clocks caused initial sampling rates to differ by several hundred Hertz; calibration reduced the mismatch to tens of Hertz but did not eliminate relative drift.The modules transmit their acoustic data over WiFi to a central computer.Source: §4.6.1 · p. 16
  • Synchronization requests sometimes failed during continuous sampling and wireless transmission, while missing or corrupted packets led to discarded recording segments and phase inconsistencies.These interruptions produced fragmented multichannel recordings.Source: §4.6.1 · p. 16, Figure 10 · p. 16

4.6 Practical Sensing Challenges

  • Coarse packet alignment and initial sampling-rate calibration are insufficient for long-term phase coherence; reliable TDOA requires synchronized hardware or continuous clock-offset and drift compensation.Practical deployments also need packet-loss detection, phase calibration, accurate mobile-receiver positioning, and multipath-aware measurement processing.Source: §4.6.1 · p. 16, §4.6.2 · p. 17

4.6.2 Multipath-Induced Phase Instability

  • Measured TDOA fluctuates substantially over one second despite fixed source-receiver geometry, with multipath, residual clock misalignment, and data loss among the possible contributors.The figure compares measured TDOA with the geometrically expected value.Source: §4.6.2 · p. 16, Figure 11 · p. 16
  • In a 20×20×20 m simulation with a 1 kHz source, first-order reflections at low absorption produced mean phase errors exceeding 40°, equivalent to about 3.8 cm of range-difference error.The reported phase distortion can affect narrowband localization even when receiver geometry is accurately known.Source: §4.6.2 · p. 16

5 Conclusion

The report concludes that mobile anchor constraints and quality-adaptive tracking can localize and track narrowband acoustic sources with sparse receivers, while practical performance also depends on reliable measurements. Simulations, robotic integration, and open-water experiments support feasibility, alongside deployment limitations from synchronization, data loss, and multipath.

5 Conclusion

  • Open-water experiments validated the complete system with sub-meter tracking error.The experiments also revealed limitations from independently synchronized receivers, wireless data loss, and multipath-induced phase instability.Source: §5 · p. 17
  • The two-stage framework uses short-baseline anchor constraints to resolve phase ambiguity during initialization, then particle-filter tracking with quality-adaptive measurement weighting.The mobile sensing platform augments sparsely distributed receivers.Source: §5 · p. 17
  • Two anchor constraints are essential for reliable 3D localization, and performance is more sensitive to anchor-pair measurement accuracy than to fixed-pair measurements.Quality-adaptive weighting improved robustness when individual receiver-channel signal quality was degraded.Source: §5 · p. 17
  • ROS-Gazebo integration demonstrated scalable surveillance operation across floating and submerged facilities.Future efforts prioritize synchronization, phase calibration, reliable multichannel acquisition, accurate mobile-receiver positioning, and multipath-aware processing.Source: §5 · p. 17
Loading 2610.34393v1…