Trojan Horse Hunt in Time Series Forecasting for Space Operations

Krzysztof Kotowski, Ramez Shendy, Jakub Nalepa, Przemyslaw Biecek, Piotr Wilczynski, Agata Kaczmarek, Dawid Pludowski, Artur Janicki, Evridiki Vasileia NtagiouarXiv 2025 · DOI · arXiv

Abstract

As the space industry increasingly relies on AI for autonomous satellite operations, the security of these models becomes critical. This paper introduces a novel Kaggle competition focused on detecting and reconstructing "Trojan" backdoor triggers hidden within time series forecasting models. By challenging researchers to reverse-engineer 45 unique triggers from poisoned neural networks, the authors aim to harden AI systems against adversarial data poisoning in high-stakes, safety-critical space environments.

Introduction: The Vulnerability of Space AI

Satellite telemetry models must be frequently retrained or fine-tuned to account for changing mission phases and environmental shifts. This creates a massive "attack surface" where an adversary could inject malicious data (poisoning) to create a backdoor. In these "Trojan Horse" attacks, a model behaves perfectly on normal data but produces a specific, harmful output when it encounters a "trigger"—a specific pattern in the input. While backdoor detection is well-studied in computer vision, it remains largely unexplored in multivariate time series forecasting.

The Threat Model: Backdoors in Forecasting

Unlike classification tasks where a trigger might change a label (e.g., a stop sign misclassified as a speed limit), in forecasting, the trigger forces the model to predict a specific "abnormal" future sequence. This could trick operators into thinking a satellite is failing, potentially triggering unnecessary and costly "safe mode" cycles.

Visualization of a poisoned input and its manipulated forecast compared to clean data.

The paper defines the poisoned segment as $segment_{poisoned} = segment_{clean} + trigger$. This additive property is what participants must exploit to reconstruct the original signal. As shown in the telemetry visualization, the poisoned input (red) deviates significantly from the clean input (green), leading to a drastically different forecast.

Methodology: Reverse-Engineering the Trigger

The competition task is essentially a reverse-engineering challenge. Participants are given 45 "poisoned" versions of the N-HiTS (Neural Hierarchical Interpolation for Time Series) model. The core technical components include trigger injection and a baseline defense adapted from "Neural Cleanse."

The workflow for trigger injection, model poisoning, and the subsequent reverse engineering process.

The authors adapted the Neural Cleanse method by using an optimization function to find a candidate trigger ($\delta$) that maximizes the difference between forecasts on clean vs. triggered data, while ensuring the forecast "tracks" the shape of the input. The loss function balances divergence, tracking, and energy (maximizing the $L_2$ norm of the trigger).

Dataset and Evaluation

The competition utilizes the ESA-ADB (European Space Agency Benchmark for Anomaly Detection in Satellite Telemetry), featuring real-life multivariate telemetry from three large ESA missions. For this challenge, the focus is on a subset of 3 specific channels (channels 44-46) from "Mission 1."

To evaluate submissions, the authors introduced NMAE_range (Normalized Mean Absolute Error bounded by range). This metric is scale-invariant and robust to outliers, ensuring that the competition is won by the most precise reconstruction rather than the one that gets lucky on high-amplitude segments.

Conclusion and Future Work

The current framework assumes triggers are additive and of a known length (75 samples). Future work will explore more complex, non-linear, or "global" triggers that are harder to isolate. While focused on space, these methods are applicable to other safety-critical domains like power grid management or medical monitoring. The competition, hosted on Kaggle through August 2025, provides a sandbox for developing the next generation of secure AI for space operations.

Summary updated

BibTeX
@misc{kotowski2025trojan,
  author = {Krzysztof Kotowski and Ramez Shendy and Jakub Nalepa and Przemyslaw Biecek and Piotr Wilczynski and Agata Kaczmarek and Dawid Pludowski and Artur Janicki and Evridiki Vasileia Ntagiou},
  title = {{Trojan Horse Hunt in Time Series Forecasting for Space Operations}},
  year = {2025},
  eprint = {2506.01849},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.01849},
}
Figure