When appearance and motion disagree: evaluating class--motion consistency against disguised uncrewed surface vessels

Jon Reifschneider

TL;DR

This paper introduces a "class–motion consistency" defense that detects disguised uncrewed surface vessels (USVs) by verifying if their radar-measured movement matches their camera-assigned vessel type. Using a novel "hybrid replay" evaluation protocol, the authors demonstrate that simple speed and closing-rate thresholds can flag up to 95% of simulated attacks while imposing a significant "time-to-target" penalty on attackers trying to remain stealthy.

What Problem Does This Paper Address?

The rise of USVs in maritime conflicts (e.g., the Black Sea and Red Sea) has created a high-stakes "needle in a haystack" problem. Hostile craft often disguise themselves as benign civilian vessels—like fishing boats or recreational yachts—to blend into busy port traffic.

Current defenses typically rely on two channels that can be independently defeated:

  1. Cameras: Computer vision systems can be fooled by physical disguises (dummies on deck) or adversarial hull patches that suppress detection.
  2. Radar: While radar provides precise motion data, it cannot identify a vessel's class.

The authors address the "residual threat" where an attacker successfully achieves a benign camera label (e.g., "fishing boat") but maintains the high-speed motion of a weaponized craft.

Figure 1: The attacker can change the craft’s appearance and choose its physical motion but cannot falsify the radar measurement. Solid arrows are runtime data flows and the dashed path at the bottom is the benign vessel motion model, fit offline on historical vessel tracks.

Methods

The core contribution is a cross-modal consistency check. The system takes a vessel class assigned by a camera and compares it against a motion model derived from historical radar/AIS data for that specific class.

Key technical components include:

  • Hybrid Replay Protocol: Since real USV attack data is scarce, the authors developed a protocol that takes real-world benign vessel tracks (prefixes) and "splices" in simulated attack maneuvers (run-ins) at various speeds and geometries (direct, dog-leg, and constant-angle).
  • Threshold Checks: Instead of complex black-box models, the authors focus on two interpretable features:
  • Speed-cutoff: Flags vessels exceeding the 95th percentile speed for their claimed class.
  • Asset-relative closing rate: Flags vessels approaching a protected asset faster than typical for their class.
  • Adversarial Evaluation: They tested the camera's robustness against digital adversarial patches using white-box, gray-box, and black-box access models.
Table 1: Closest lines of prior work, by viewpoint, evidence used, and defense. Among the cited studies, none pairs a shore-based view of an incoming craft with a defense that uses world-coordinate motion learned from real benign traffic.

Datasets

The study utilizes two primary, non-paired datasets:

  • Image Dataset: A merged set of 25,262 images containing 71,501 labeled vessels from public sources (SeaShips, ABOShips, etc.) plus a curated set of USV images.
  • Track Dataset: One week of MarineCadastre AIS data from the U.S. coast, consisting of over 153,000 voyage tracks from 30,053 vessels. This was used to build the "normal" motion models for various vessel classes.

Key Insights and Ideas

  • Patches vs. Disguise: Adversarial patches were found to mostly suppress detection (making the boat "invisible" to the camera) rather than cause misclassification. However, suppression is easily caught by a "presence check" (radar sees a boat, camera doesn't). Physical disguise is the more dangerous threat because it creates a "matched" but incorrect detection.
  • Geometric Complementarity: The two motion checks cover each other's blind spots. A direct run-in is easily caught by the closing-rate check. An attacker trying to lower their closing rate by steering off-axis (oblique approach) remains visible to the track-only speed check.
  • The "Stealth Tax": To avoid both checks, an attacker must move slowly. The authors show that staying below these thresholds extends a 6-nautical-mile approach to at least 31 minutes, giving defenders significantly more time to intervene via other means.
  • Simplicity Wins: Simple, class-conditioned thresholds outperformed more complex multivariate scoring methods like Gaussian Mixture Models or Isolation Forests on direct run-ins.

Key Results

  • Detection Performance: For recreational-labeled vessels, the closing-rate check flagged 94–95% of direct attacks at 15–20 knots.
  • False Alarms: The combined defense had a 4.9% overall false-alarm rate on held-out benign traffic. This rate was higher (12.3%) for recreational traffic, which is naturally more "erratic" than cargo ships.
  • Adversarial Robustness: While patches suppressed detection in up to 29% of attempts, they never successfully caused a "hostile-to-benign" label change in 420 tested configurations.
  • Warning Time: At class-derived speeds, the checks provided a median warning time of 1.1 to 2.3 minutes for close-in commits, extending to nearly 20 minutes for detections at the 12nm range.

Limitations and Future Work

  • AIS as a Proxy: The study uses AIS data as a stand-in for radar. Real radar data involves noise, fragmentation, and association errors that weren't modeled here.
  • Synthetic Motion: The simulated attack run-ins use constant speeds and instantaneous turns; real-world dynamics (acceleration, sea-state effects) are more complex.
  • Small Samples: The samples for fishing and sailing vessels were relatively small, making class-specific conclusions for those categories less robust.
  • Future Directions: The authors suggest moving to physical trials and evaluating the system's performance when integrated with human operators to measure "alarm fatigue."

Summary updated

Abstract

Defense systems for maritime assets must distinguish approaching benign craft from uncrewed surface vessels (USVs) with sufficient warning to act. Shore cameras using computer vision can separate hostile vessels from benign traffic, but these systems may be attacked using an adversarial hull patch or physical disguise. We evaluate adversarial hull patches and find they suppress detection in up to 29% of attempts but never cause misclassification as benign. Physical disguise, however, can bypass both a camera check and a radar presence check. To evaluate defenses, we develop a hybrid replay protocol from automatic identification system (AIS) tracks that combines recorded benign motion with direct or off-axis attack run-ins. We find that for recreational-labeled vessels attacking via direct run-in within 4 nautical miles, an asset-relative closing-rate check flags 94-95% at 15-20 knots, compared with 12% for a track-only speed check. The checks are complementary on off-axis approaches, constraining the attacker to remain below both thresholds and extending the time required to reach the asset. Flagging on either check has a 4.9% held-out false-alarm rate overall and 12.3% on recreational traffic. The replay-based evaluation protocol supports comparative test and evaluation of defense methods.
BibTeX
@misc{reifschneider2026when,
  author = {Jon Reifschneider},
  title = {{When appearance and motion disagree: evaluating class--motion consistency against disguised uncrewed surface vessels}},
  year = {2026},
  eprint = {2611.00367},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00367},
}
Figure