AIS was built for safety, not security. It broadcasts what a vessel says it is — a self-declaration, on the honour system. The result is a surveillance layer that can be gamed three ways: switch it off and go dark, transmit a false identity or position, or jam/spoof the GNSS underneath so the ship itself doesn't know where it is. The last one is the nasty case — the vessel reports its position faithfully, and the position is still a lie. Established
Key findings
- ML detection works: Isolation-Forest spoofing identification reports 88–93% accuracy; deep models push tampered-trajectory recognition higher on scarce data. Research demonstration
- Wide-area detection at scale is real — a 966M-message pipeline found 17 spoofing + 343 jamming clusters off Korea while cutting false alarms 98.6%. Research demonstration
- The hard part is false alarms — communication-layer defects (duplicated MMSIs, stale retransmissions, timestamp errors) look exactly like spoofing unless filtered first. Research demonstration
- AIS-only evidence can't attribute — a candidate event is "consistent with spoofing," not proof of it; honest systems say so. Research demonstration
- Recurring spoofing geographies are documented — Baltic, Black Sea, Murmansk, Haifa areas show repeated signatures in global analysis. Research demonstration
Three different attacks, three different detectors
MMSI/identity spoofing — a ship transmits another vessel's identity, or a fake one. Detected by kinematic impossibility: a track that can't be one continuous voyage splits into regular + spoofed segments. The Isolation-Forest study mines exactly that — speed and distance discontinuities between adjacent points — reaching 88–93% depending on the segmentation window. Research demonstration
Trajectory tampering — a real track altered after the fact, or a fabricated track injected. Deep models like MConLSTM learn the statistics of genuine motion and flag tracks that lack the right kind of error — spoofed tracks are oftentoo clean, missing the measurement noise real GNSS always has. Research demonstration
GNSS spoofing/jamming — the ship's own receiver is deceived, so the AIS position it faithfully reports is wrong. Detection moves from single-vessel kinematics to multi-vessel coherence: one ship jumping is a glitch; a whole cluster of ships simultaneously displaced to the same impossible region is a spoofing event. That's what the wide-area pipelines exploit. Research demonstration
The false-alarm problem is the real problem
Read the wide-area study carefully: raw AIS is riddled with communication-layer defects — duplicated MMSIs, timestamp errors, stale retransmissions, multi-station rebroadcast delays — that mimic spoofing and jamming almost perfectly. Run naive anomaly detection on it and you don't find attacks; you find noise and drown the watch floor in alerts. Research demonstration
The pipeline that worked spent most of its intelligence on pre-filtering: rule-based diagnostics to discard communication faults first, then kinematic-consistency filtering, then spatiotemporal clustering requiring multi-vessel agreement before calling an event. The result — a 98.6% false-alarm reduction versus naive clustering — is the honest engineering: detection accuracy is less important than alarm quality, because a spoofing alarm that cries wolf gets ignored like every other alarm. Maritime-GPT analysis
Maritime-GPT synthesis of the wide-area detection pipeline. Maritime-GPT analysis
What we know / don't / being tested
What we know
- ML and filter-based detection catches spoofed identity, tampered tracks and GNSS events in research data.
- Multi-vessel coherence is the discriminator between attack and artifact.
- Real recurring spoofing regions are documented at scale.
What we don't know
- Attribution — AIS-only evidence flags events, it doesn't identify the attacker.
- Adaptive adversaries — detection rates against spoofers who know the detector.
- Shipboard real-time performance — most results are shore-side, retrospective analyses.
Being tested
- Wide-area GNSS-interference detection from aggregated AIS.
- Self-supervised trajectory models needing little labelled data.
- Cross-sensor confirmation — fusing AIS anomalies with radar/SAR to firm up attribution.
Maritime-GPT analysis
The useful product is a flag, not a verdict. AIS-only detection can say "this cluster is consistent with GNSS interference" — it cannot and should not say who did it or even that it definitely happened. Systems that overclaim get trusted once and ignored after the first false alarm; the honest deployment surfaces a well-filtered candidate and hands the attribution to a human with other sensors. For a vessel, the practical value is simpler: knowing your own position might be wrong — which is a bridge-safety issue before it's a security one. Maritime-GPT analysis
Open research problems
- Attribution beyond AIS — correlating anomalies with radar, SAR and RF sensing to identify the interferer.
- Shipboard self-protection — a receiver that detects it's being spoofed and tells the OOW, not just a shore analyst.
- Adaptive-adversary robustness — detection that survives a spoofer who mimics genuine kinematics.
- Threshold calibration under traffic density — alarm rates in congested vs sparse waters.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | Moderate | Strong detection results; preprints and single-region studies dominate. |
| Research maturity | Field trial | Demonstrated at scale retrospectively; shipboard real-time unproven. |
| Operational relevance | High | GNSS interference is a live, documented operational threat. |
| Confidence in conclusions | High | Numbers per-study; attribution limits stated by the sources themselves. |
Related tools & reading
Sources
- Identification of spoofing ships from AIS data via trajectory segmentation and Isolation Forest — J. Marine Science and Engineering 11(8):1516Peer-reviewed
Isolation-forest detection of MMSI-spoofing vessels: 88–93% accuracy depending on trajectory-segmentation window — classical ML as the baseline method. - A deep-learning approach for identifying intentional AIS signal tampering in maritime trajectories — J. Marine Science and Engineering 13(4):660Peer-reviewed
MConLSTM self-supervised model for tampered-trajectory recognition; reports ~23% accuracy gain over a general model with scarce training data. - Wide-area GNSS spoofing and jamming detection from AIS — arXiv 2603.11055 (2026)Preprint
966M AIS messages off Korea: rule-filter + IMM + DBSCAN pipeline detects 17 spoofing and 343 jamming clusters, cutting false alarms 98.6% vs naive clustering. - SeaSpoofFinder — potential GNSS spoofing event detection using AIS — arXiv 2602.16257 (2026)Preprint
Global AIS pipeline revealing recurring spoofing signatures in the Baltic, Black Sea, Murmansk and Haifa areas — and a candid account of AIS-only limits.
Detection rates and event counts are the cited studies' results on their stated data. Accessed 5 October 2026. Editorial analysis only — not security advice.