Sequential RF Fingerprinting for AIS: Rejecting Unknown Vessels and Insider MMSI Spoofs Days After Enrollment

summary

Video file (mp4)

The gist

Any AIS transmitter can claim any vessel’s identity (MMSI), because nothing in the protocol authenticates it; only the transmitter’s analog hardware reveals who is really sending page 1.

In short

The system uses sequential RF fingerprinting to identify vessels by analyzing physical layer signals. It rejects unknown vessels after ten days and detects insider MMSI spoofs with high accuracy by comparing incoming signals against enrolled vessel fingerprints. This method relies on accumulating evidence over time and dynamically adjusting detection thresholds to handle signal drift.

Key concepts

Sequential RF Fingerprinting
This technique analyzes the physical radio signals (I and Q channels) of an AIS burst sequentially. It uses a frozen machine learning model trained only on known vessels to generate a unique 'fingerprint' for each transmission. This allows the system to detect deviations from known patterns over time.
GEM-CUSUM Statistic
This is a statistical tool used to accumulate evidence from successive signal bursts. It measures the distance of new signals from the enrolled fingerprints in a feature space, using geometric entropy minimization. By accumulating these distances, it builds a robust score that detects subtle changes or unknown signals effectively.
Claimed-MMSI Verification
This test specifically checks if an incoming signal claiming a certain MMSI matches the actual physical fingerprint of that vessel. It scores the burst against the claimed identity's features. This is highly effective at exposing insider spoofs because it verifies not just the transmitter, but also whether it is authentically mimicking a specific enrolled vessel.
Threshold Refresh Rule
The system dynamically updates its detection threshold using only genuine traffic data. This prevents the detector from becoming overly sensitive or too insensitive as real signals naturally drift over time. Maintaining this rule keeps the system's false alarm rate stable despite changes in the deployed environment.

Terminology used across episodes

This episode discusses

The paper

Sequential RF Fingerprinting for AIS: Rejecting Unknown Vessels and Insider MMSI Spoofs Days After Enrollment · Read on arXiv

Fawaz Abdul Razak, Bora Turgut, Mason Clark, Yasin Yilmaz

University of South Florida

Any AIS transmitter can claim any vessel's identity (MMSI), because nothing in the protocol authenticates it; only the transmitter's analog hardware reveals who is really sending. We treat AIS traffic as streams of bursts keyed by claimed MMSI and accumulate physical-layer evidence with a sequential test. Each burst is scored by its k-nearest-neighbor distance to enrolled fingerprints in the feature space of a classifier whose weights are fixed after training, calibrated into a tail probability, and accumulated by a geometric-entropy-minimization cumulative-sum (GEM-CUSUM) statistic. The same test rejects unknown transmitters when scored against all enrolled vessels, and exposes insider spoofs when scored against the claimed vessel. A threshold refreshed from genuine traffic alone keeps it at a stated false-alarm rate as fingerprints drift. On AIS in-phase/quadrature (I/Q) recordings captured over 13 days at the Istanbul Strait (Bosphorus), the detector rejects unseen vessels at least 10 days after enrollment. Evaluated over 50-burst trials with every vessel weighted equally, it reaches an area under the ROC curve (AUROC) of 0.956 (95% confidence interval 0.913-0.989), where the best score computed from one burst at a time reaches 0.899. A threshold refreshed from the first quarter of the enrolled vessels' traffic holds 5.3% false alarms (3.0-9.2% across five independently trained encoders) at 88.8% detection. Scored against the claimed MMSI, enrolled transmitters spoofing another vessel are rejected at 0.991, a threat that open-set tests cannot see (AUROC 0.500, chance level). The lightweight Transformer GLFormer, reimplemented from its paper (no code is public) and retrained on the same data, reaches 0.652 in the same detector with its published in-phase input and 0.849 when given the same I/Q input as our encoder.

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Sequential RF Fingerprinting for AIS".

Dev: Here's the gist of the paper. Any AIS transmitter can claim any vessel’s identity (MMSI), because nothing in the protocol authenticates it; only the transmitter’s analog hardware reveals who is really sending page 1.

Rosa: The detector rejects unseen vessels at least 10 days after enrollment page 3, and exposes insider spoofs when scored against the claimed vessel page 3.

Dev: Pruning redundant reasoning tokens The paper introduces a sequential RF fingerprinting method for Automatic Identification System (AIS) traffic that uses physical-layer evidence to reject unknown vessels and verify insider MMSI spoofs page 8.

Rosa: This detector scores each burst by its k-nearest-neighbor distance to enrolled fingerprints in a frozen classifier's feature space, accumulates this evidence using a geometric-entropy-minimization cumulative sum (GEM-CUSUM) statistic, and employs a threshold refreshed from genuine traffic alone page 8.

Title and authors: Rosa: So, let's talk about the title again, "Sequential RF Fingerprinting for AIS: Rejecting Unknown Vessels and Insider MMSI Spoofs Days After Enrollment." It really lays out the two main things this paper is trying to achieve >

Dev: Basically, it’s about moving beyond just checking the claimed MMSI number because that doesn't prove anything; it focuses on using the physical layer characteristics of the radio bursts themselves >

Rosa: Right. They are treating AIS traffic as streams of bursts keyed by claimed MMSI and accumulating physical-layer evidence with a sequential test, which is what they call GEM-CUSUM >

Dev: The implication here is that you can establish a kind of long-term memory for vessel fingerprints, allowing the system to reject unknown transmitters even if they look similar to enrolled ones later on >

Taro: I wonder how robust this sequential accumulation is when the underlying physical signals themselves are drifting over time, which is something they mention in their limitations >

Rosa: They address that by using a threshold refreshed from genuine traffic alone, which keeps the detector at a stated false alarm rate even as those fingerprints drift >

Dev: That’s important because if you have a fixed threshold, it just gets worse and worse over time as the real traffic patterns change >

Taro: So they are essentially making the detection process adaptive to how much things have changed since enrollment >

Rosa: And they also test this on an open-set test against unknown transmitters, and that seems to be where the long-term rejection capability comes from >

The paper's summary: Dev: Moving into the summary, they detail how the system actually works on a technical level. It uses a frozen feature encoder, specifically a one dimensional ResNet eighteen trained only on enrolled vessels > <ref:2610.19115#pg2>

Rosa: And those features are derived from the I and Q channels of the received burst; crucially, they state that the MMSI field never enters into the input to that encoder >

Dev: That means it’s purely analyzing the analog hardware's transmission properties, which is what they claim reveals who is really sending page four >

Taro: I'm thinking about how training only on enrolled vessels impacts its ability to generalize when an entirely new class of transmitters appears >

Rosa: They calibrate those classifier weights into a tail probability using a held-out calibration set, which helps define what constitutes a significant deviation from normal traffic >

Dev: And then they use this to build the GEM-CUSUM statistic, which is essentially accumulating evidence based on how far each burst is from its k nearest neighbors in that feature space >

Taro: So it’s not just one shot; it’s a running tally of evidence over time, which sounds like the mechanism for rejecting those unseen vessels after a period of observation >

Rosa: It really means they are building a cumulative statistic over time, and this is what allows them to reject unknown transmitters at least ten days after enrollment > <ref:2610.19115#pg1>

The paper's improvements: Dev: The improvements section highlights that they formulate the detector as two sequential tests: an open-set test against unknown transmitters and a claimed-identity test against insider spoofs >

Rosa: And they emphasize that the GEM-CUSUM statistic accumulates loglikelihood ratio evidence, where each burst contributes a statistic based on the tail probability of calibration distances >

Taro: So, what’s the practical difference between these two tests when you're trying to catch different kinds of bad actors >

Rosa: The open-set test is designed to reject unseen hardware that claims the enrolled vessel it most resembles, while the claimed-identity test targets insider spoofs specifically >

Dev: And they show that the claimed-identity test rejects enrolled transmitters spoofing another vessel at a score of zero point nine nine one when scored against the claimed MMSI > <ref:2610.19115#pg1,when scored against the claimed>

Taro: That zero point nine nine one score sounds much stronger than what an open-set test scores them at, which is around zero point five zero zero in some comparisons they mention >

Rosa: Yeah, so it’s a way to verify the claimed identity against the hardware's fingerprint to catch those insiders when they try to impersonate someone >

Dev: And the threshold refresh rule is key there; it’s refreshed from genuine traffic only, which keeps the detector at a stated false alarm rate as fingerprints drift >

Conclusion: Rosa: So, summarizing this paper on "Sequential RF Fingerprinting for AIS: Rejecting Unknown Vessels and Insider MMSI Spoofs Days After Enrollment," they show that you can reject unknown transmitters after about ten days and catch insider spoofs at a score of zero point nine nine one > <ref:2610.19115#pg1,Sequential RF Fingerprinting for AIS: Rejecting Unknown Vessels and Insider MMSI Spoofs>

Dev: The main implication is that the physical hardware signature behind the signal is more reliable than the claimed identity itself, especially over time >

Taro: I just think it's neat that they made it sequential, because it doesn't rely on a single static setting forever >

Rosa: It’s about making the system adaptive to drift by using genuine traffic only for threshold refresh and monitoring how far deployed traffic has moved from enrollment >

Dev: The paper suggests that future work should test other receivers and sites, adversaries that synthesize or replay fingerprints, and look at full-frame captures that include the training sequence >

Taro: It makes sense to test against synthesized transmitters too, because an opportunistic insider is just one type of adversary they focused on >

Rosa: Exactly. We’re looking at how this sequential approach handles those different kinds of threats moving forward >

More episodes

← Home

Watch the video for this episode