Adaptive Transient Fault Model for Sensor Attack Detection

Minsu Jo, Junkil Park, Youngmi Baek, Radoslav Ivanov, James Weimer, Sang Hyuk Son, Insup LeeCPSNA 2016 · DOI

TL;DR

This paper tackles the challenge of detecting sensor attacks in systems with multiple operating modes, where a single, fixed model for sensor behavior is insufficient. They propose an "adaptive transient fault model" that adjusts its parameters based on the system's current operating mode, significantly improving attack detection accuracy compared to existing methods.

What Problem Does This Paper Address?

Cyber-Physical Systems (CPS) are increasingly reliant on multiple sensors for robust operation and safety. However, these sensors can be vulnerable to attacks, such as spoofing, where malicious data is injected into the system. While redundant sensors can help detect and mitigate these attacks, a key challenge arises in systems with varying operating modes (e.g., a vehicle operating at different speeds). Existing sensor attack detection methods often use a "transient fault model" (TFM) with fixed parameters. To ensure safety across all modes, these parameters are typically set conservatively for the worst-case scenario. This conservatism, however, leads to reduced detection effectiveness in less demanding operating modes where sensor noise is naturally lower. The paper aims to bridge this gap by developing an adaptive TFM that dynamically adjusts to the system's operating mode, thereby enhancing attack detection performance without compromising safety.

Methods

The core of the proposed solution is an adaptive transient fault model (TFM). This model builds upon the existing TFM framework, which characterizes sensor behavior using parameters like error bound ($\epsilon$) and maximum transient faults per window ($\hat{e}$). The novelty lies in making these parameters adaptive to the system's operating mode, specifically the reference speed of the vehicle.

To achieve this adaptivity, the authors introduce two key components:

  1. Lookup Table Construction: A lookup table is created that stores TFM parameter sets for various operating speeds. This table is populated by training data collected from the system.
  2. Automated Transient Fault Modeling: To efficiently generate the TFM parameters for different speeds, the paper proposes an automated method for finding the "knee point" in a plot of error bound ($\epsilon$) versus the ratio of faults to window size ($\epsilon/\omega$). This knee point represents an optimal trade-off between detection sensitivity and false alarms. The automated method calculates the perpendicular distance from each data point to a diagonal line and selects the point with the maximum distance, effectively identifying the knee point without manual intervention.

At runtime, the system uses the current reference speed to query the lookup table and select the appropriate TFM parameters for attack detection. If the exact speed is not in the table, it defaults to the parameters of the next higher speed to maintain a conservative approach and avoid false alarms.

Datasets

The paper utilizes data collected from an unmanned ground vehicle called Jackal. This data includes sensor measurements from left and right encoders, and GPS, sampled at 20 Hz and 10 Hz respectively. The attack detection is performed at the lowest sensor frequency, 10 Hz. The authors collected data by driving the Jackal robot at various constant speeds (0.4 m/s to 1.6 m/s) to simulate different operating modes. They also introduced simulated bias and random attacks on the GPS sensor for evaluation.

Key Insights and Ideas

  • Operating Mode Matters: The paper highlights that a one-size-fits-all approach to sensor attack detection is suboptimal for systems with varying operating modes. The inherent characteristics of sensor noise change with operating conditions, and a fixed TFM cannot capture this.
  • Conservatism vs. Performance Trade-off: Using overly conservative TFM parameters to ensure safety across all modes significantly degrades detection performance in less critical modes. The adaptive approach strikes a better balance.
  • Automated TFM Training is Crucial: Manually determining TFM parameters for multiple operating modes is tedious and impractical. The proposed automated method for finding the "knee point" is a significant contribution, enabling efficient generation of the adaptive lookup table.
  • Speed as a Proxy for Operating Mode: The paper uses vehicle speed as a direct indicator of the operating mode, demonstrating its effectiveness in adapting TFM parameters.

Key Results

  • Improved Attack Detection: In a motivating example involving a simulated bias attack on GPS during a speed change, the proposed adaptive system successfully detected the attack, whereas the existing system failed.
  • Effective in Random Attacks: For random attacks, the adaptive system showed significantly higher detection rates, especially for smaller attack magnitudes, compared to the existing system.
  • No False Alarms: Both the existing and proposed systems demonstrated a 0% false alarm rate in normal operation across various test scenarios, indicating that the adaptive approach does not compromise safety.
  • Magnitude Matters: Detection rates for both systems were generally higher for larger attack magnitudes. However, the proposed system showed a more consistent and improved performance across different magnitudes, particularly for random attacks.

Limitations and Future Work

  • Online Learning: The authors acknowledge that the current TFM parameter selection relies on a pre-constructed lookup table. For future work, they plan to investigate online learning schemes to find proper TFM parameters at runtime, allowing for even greater adaptability.
  • System Dynamics: Incorporating more detailed system dynamics into the attack detection system is suggested as a way to improve performance against stealthy and collusion attacks.
  • Manual vs. Automatic TFM: While the paper presents an automated method for TFM parameter selection, they validated its results against manually selected parameters, confirming its accuracy. However, the underlying principle of identifying the "knee point" is still based on a graphical interpretation.
  • Limited Attack Types: The study primarily focuses on bias and random attacks. More sophisticated attacks, especially those that exploit knowledge of the detection scheme, are not explicitly addressed in this work.

Summary updated

BibTeX
@inproceedings{jo2016adaptive,
  author = {Minsu Jo and Junkil Park and Youngmi Baek and Radoslav Ivanov and James Weimer and Sang Hyuk Son and Insup Lee},
  title = {{Adaptive Transient Fault Model for Sensor Attack Detection}},
  booktitle = {CPSNA},
  year = {2016},
  doi = {10.1109/cpsna.2016.21},
}
Figure