🤖 AI

Rarity-Gated Context Conditioning for Offline Imitation Learning-Based Maritime Anomaly Detection

The paper proposes Rarity-Gated Feature-wise Linear Modulation (RGFiLM), a novel module that dynamically adjusts context conditioning based on a data-driven rarity score to effectively reduce false alarms and improve the precision of maritime anomaly detection in rare environmental regimes.

Original authors: Yongmin Kim, ByeongHoon Jeon, Sungil Kim

Published 2026-06-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yongmin Kim, ByeongHoon Jeon, Sungil Kim

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 you are a security guard watching a busy harbor. Your job is to spot ships doing something "weird."

Here's the problem: Context matters.
If a ship suddenly swerves wildly during a calm, sunny day, that's definitely weird. But if a ship swerves wildly during a massive hurricane, that's actually normal behavior! It's just trying to stay safe.

The paper argues that most computer systems trying to spot these "weird" ships make a mistake: they treat every situation the same way. They get confused when the weather is extreme (a "rare" situation) because they haven't seen enough examples of it. This leads to two bad outcomes:

False Alarms: The system screams "Emergency!" when a ship is just reacting to a storm.
Unstable Decisions: The system flips back and forth, unsure if a behavior is normal or not.

The Solution: The "Rarity Gate"

The authors propose a new tool called RGFiLM (Rarity-Gated Feature-wise Linear Modulation). To understand how it works, let's use an analogy.

Think of the computer system as a chef trying to cook a dish (detecting anomalies).

The Ingredients (Context): The weather, wind, and waves are the ingredients.
The Recipe (The Model): The standard recipe says, "If the ship moves fast, it's normal."
The Problem: The chef has cooked this dish a thousand times on sunny days (common context) but only once during a hurricane (rare context). When a hurricane hits, the chef panics and might ruin the dish or over-season it because they aren't used to it.

RGFiLM adds a "Rarity Gate" to the chef's apron.

Checking the Rarity Score: Before cooking, the system checks a "Rarity Meter." It asks, "How often do we see this specific weather?"
- If the weather is common (sunny day), the meter says, "We've seen this a million times. Stick to the standard recipe, but be careful not to overreact."
- If the weather is rare (extreme storm), the meter says, "Whoa, this is unusual! We need to pay extra attention to how the weather changes the rules."
The Gate Opens or Closes:
- In Common Weather: The gate stays mostly closed. The system relies on its standard, safe understanding of how ships behave. It doesn't let the weather change the rules too much, preventing it from getting confused by minor variations.
- In Rare Weather: The gate swings wide open. The system says, "Okay, this is a special case. Let the weather heavily influence our decision." It allows the system to say, "Ah, that crazy swerve is actually normal because of this rare storm," preventing a false alarm.

How They Tested It

The researchers tested this on maritime ships using real data:

The Data: They used ship tracking data (AIS) and weather data (ERA5).
The Test: They took normal ship paths and artificially added a "detour" (a fake weird turn) to see if the system could spot it.
The Result:
- Old systems (without the gate) sounded the alarm too often, especially when the weather was weird.
- The new RGFiLM system was much better. It caught the fake detours just as well as the old systems but sounded far fewer false alarms when the weather was extreme.

The Bottom Line

The paper claims that by teaching the computer to recognize how rare a situation is, we can make it smarter. Instead of blindly reacting to every unusual weather pattern, the system learns to be conservative when things are normal and decisive when things are rare. This stops the system from crying "Wolf!" every time a storm rolls in, while still catching the actual bad guys.

In short: It's a smarter security guard that knows when to ignore the noise of a storm and when to actually look closely at the ship.

Technical Summary: Rarity-Gated Context Conditioning for Offline Imitation Learning-Based Maritime Anomaly Detection

Problem Statement

Contextual anomaly detection (CAD) aims to identify abnormal behavior $b$ conditional on context variables $c$ (e.g., environmental conditions). However, practical deployments often suffer from frequency bias: context distributions are highly imbalanced, with routine contexts being abundant and rare contexts (e.g., severe weather, extreme operating regimes) being scarce.

Under such imbalance, standard context-conditioned models tend to mishandle rare regimes. Because these regimes are underrepresented in training data, models may produce unstable decisions or excessive false alarms when encountering them. Existing mitigation strategies, such as "backing off" to a context-agnostic baseline when local evidence is sparse, effectively down-weight rare contexts. This is counterproductive, as rare contexts are often safety-critical and require careful, rather than diminished, contextual influence.

Methodology: Rarity-Gated FiLM (RGFiLM)

The authors propose Rarity-Gated Feature-wise Linear Modulation (RGFiLM), a modular conditioning module designed to regulate how strongly context influences intermediate representations based on the rarity of the current environment. The method is integrated into a sequential anomaly scoring pipeline, specifically instantiated within OIL-AD (Offline Imitation Learning-based Anomaly Detection).

1. Core Architecture:
The approach operates on trajectories $\tau = \{(s_t, e_t, a_t)\}_{t=1}^T$ , where $s_t$ is the state, $e_t$ is the exogenous context (environment), and $a_t$ is the action.

State Representation: A sequence encoder produces a latent state $h_t = f_{state}(s_{1:t})$ .
Context Modulation (FiLM): An environment encoder maps $e_t$ to scaling ( $\gamma_t$ ) and shifting ( $\beta_t$ ) parameters, creating a context-modulated representation $\tilde{h}_t = \gamma_t \odot h_t + \beta_t$ .
Rarity Extraction: A rarity score $r_t \in [0, 1]$ is estimated from the empirical distribution of a subset of environmental variables (e.g., wave height, wind speed). The authors use the Mahalanobis distance of the environmental vector from the training mean, mapped through a Chi-squared cumulative distribution function (CDF).
Rarity Gate: The final representation $h^*_t$ $h_{t}^{*}$ is a weighted interpolation between the original state $h_t$ $h_{t}$ and the modulated state $\tilde{h}_t$ $\tilde{h}_{t}$ :
$h^*_t = w_t \tilde{h}_t + (1 - w_t) h_t$
The gating weight $w_t$ $w_{t}$ is derived from a small neural network $g$ $g$ that takes $[e_t, r_t]$ $[e_{t}, r_{t}]$ as input. Crucially, the network output is scaled by a temperature factor dependent on $r_t$ $r_{t}$ : $\tilde{s}_t = s_t(1 + \kappa r_t)/\tau$ $\tilde{s}_{t} = s_{t} (1 + κ r_{t}) / τ$ .
- Behavior: Under rare contexts (high $r_t$ ), the gate becomes "decisive" (sharper sigmoid), forcing a strong preference for either the modulated or original representation. Under frequent contexts (low $r_t$ ), the gate remains "conservative" (softer mixture), preventing the model from relying too heavily on extreme context effects where data is abundant.

2. Integration:
RGFiLM replaces the standard state representation in the OIL-AD framework. The model is trained on normal trajectories to learn a policy $\pi(a_t | s_t, e_t)$ . Anomalies are detected as deviations in decision sequences (low likelihood of observed actions given the rarity-calibrated state).

Key Contributions

Identification of Frequency Bias: The paper explicitly identifies that treating all contexts symmetrically in imbalanced distributions leads to false alarms in rare, safety-critical regimes.
Rarity-Aware Conditioning Mechanism: The proposal of RGFiLM, which interpolates between context-agnostic and context-modulated representations using a rarity-dependent gate. Unlike prior "back-off" methods, RGFiLM does not down-weight rare contexts but rather calibrates the strength of their influence.
Empirical Validation: Demonstration of the method in a maritime trajectory anomaly detection setting using AIS motion sequences and ERA5 environmental data.

Experimental Results

The method was evaluated on a detour scenario in the Bass Strait region (Australia), where synthetic lateral detours were injected into normal trajectories. The goal was to detect these deviations while distinguishing them from environment-induced route changes.

Performance Metrics: Evaluated over 20 runs using Precision, Recall, F1-score, and False Positive Rate (FPR).
Comparisons: RGFiLM was compared against:
- Context-agnostic OIL-AD.
- OIL-AD with naive concatenation.
- OIL-AD with standard FiLM.
- OIL-AD with a learnable gate but without rarity awareness.
- External baselines: ROCOD and CAD for MTS.
Findings:
- RGFiLM achieved the best mean F1–FPR trade-off, obtaining the highest mean F1 (0.595) and the lowest mean FPR (0.097).
- Compared to the standard FiLM baseline, RGFiLM improved F1 (0.577 $\to$ 0.595) and Recall while maintaining a comparable FPR.
- Compared to the "Gated FiLM (no rarity)" ablation, RGFiLM significantly reduced FPR (0.108 $\to$ 0.097) while improving F1, isolating the benefit of the rarity score.
- Rarity Stratification: When analyzing performance by environmental rarity (top 20% rare vs. bottom 80% frequent), RGFiLM maintained the highest F1 and lowest FPR in both regimes. The performance gain over standard FiLM was more pronounced in the rare regime.

Significance and Claims

The paper claims that explicitly accounting for context rarity is an effective approach for reducing false alarms in context-sensitive anomaly detection. By making the modulation strength dependent on a data-driven rarity score, RGFiLM mitigates frequency bias without discarding critical information from rare regimes.

The authors note that their evaluation relies on synthetic detour injection with fixed magnitude, which may not cover the full diversity of real-world anomalies. Additionally, the rarity estimation depends on the empirical context distribution and could be sensitive to domain shifts. Future work is suggested to extend the approach to broader anomaly types and operational periods. The study concludes that rarity-aware conditioning improves robustness in sequential decision-based anomaly detection, particularly in safety-critical maritime environments.