Multi-IF : An Approach to Anomaly Detection in Self-Driving Systems

Kun Cheng, Yuebin Bai, Yuan Zhou, Chao Yu, Yang LiuarXiv 2020 · OpenReview

Abstract

Autonomous driving vehicles (ADVs) are implemented with rich software functions and equipped with many sensors, which in turn brings a broad attack surface. Moreover, the execution environment of ADVs is often open and complex. Hence, ADVs are always at risk of safety and security threats. This paper proposes a fast method called Multi-IF, using multiple invocation features of system calls to detect anomalies in self-driving systems. Since self-driving functions take most of the computation resources and upgrade frequently, Multi-IF is designed to work under such resource constraints and support frequent updates. Given the collected sequences of system calls, the combination of different syntax patterns is used to analyze and construct feature vectors of those sequences. By taking the feature vectors as inputs, a one-class support vector machine (OCSVM) is adopted to determine whether the current sequence of system calls is abnormal, which is trained with the feature vectors from the normal sequences. The evaluations on both simulated and real data prove that the proposed method is effective in identifying abnormal behavior after minutes of feature extraction and training. Further comparisons with existing methods on the ADFA-LD dataset also validate that the proposed approach achieves a higher accuracy with less time overhead.

Introduction

Autonomous Driving Vehicles (ADVs) are essentially high-performance computing platforms on wheels. With software stacks like Autoware exceeding 270,000 lines of code, these systems manage a massive array of sensors and network interfaces. This complexity creates a significant attack surface, ranging from malicious software updates to remote exploits via Bluetooth or on-board entertainment systems.

Traditional Intrusion Detection Systems (IDS) often fail in the ADV context due to three primary constraints:

  1. Resource Scarcity: Most onboard compute is dedicated to safety-critical driving functions (sensing, planning, and acting), leaving minimal overhead for security monitoring.
  2. Frequent Updates: Modern ADVs receive frequent over-the-air (OTA) updates. Security models that require weeks to train are obsolete by the time they are deployed.
  3. Non-deterministic Behavior: The dynamic nature of driving environments and the use of machine learning in perception mean that "hard-coded" security rules are often too brittle.

The Multi-IF Methodology

The Multi-IF approach focuses on system calls (syscalls)—the fundamental requests a program makes to the operating system kernel. By monitoring these calls, Multi-IF can detect deviations in program execution that signal an intrusion.

Feature Extraction and Pattern Discovery

Multi-IF utilizes a pipeline based on $k$-grams, which breaks syscall sequences into short snippets of length $k$. The "Multi" in Multi-IF refers to the combination of multiple pattern lengths (e.g., combining patterns of length 1 and length 8). This multi-feature synergy allows the system to capture both fine-grained and structural behavior of the software.

One-Class SVM (OCSVM)

Because malicious data is rare in real-world driving scenarios, the system employs a One-Class Support Vector Machine. The OCSVM is trained exclusively on "normal" execution data, learning the boundaries of legitimate behavior. Any sequence falling outside these boundaries is flagged as an anomaly.

Heuristic Optimization

To maintain efficiency, the researchers introduced a three-step heuristic to identify the optimal combination of pattern lengths. This ensures the highest detection accuracy without unnecessary computational bloat.

Experimental Evaluation

The researchers evaluated Multi-IF across three distinct datasets to test its versatility and performance:

  1. GNSS Localization Partition: Real-world data from an "ACRONIS" vehicle (a modified Toyota COMS) running Autoware.
  2. Mapping Partition: Data gathered during field tests in Singapore.
  3. ADFA-LD: A standard public benchmark for host-based anomaly detection used to compare Multi-IF against state-of-the-art semantic models.

Key Results

Multi-IF demonstrated high efficacy and industry-leading efficiency:

  • Detection Performance: On the Localization and Mapping tasks, the system achieved a 100% Detection Rate (DR) with a False Alarm Rate (FAR) of less than 0.7%.
  • Benchmark Comparison: On the ADFA-LD dataset, Multi-IF reached an 83.6% DR with an 18.9% FAR, outperforming several existing methods.
  • Computational Efficiency: While competing semantic-based models can take weeks to train, Multi-IF completed feature extraction for the localization task in approximately 7 minutes, with SVM training taking less than one second.

Discussion and Future Work

The primary advantage of Multi-IF is its reliance on "syntax" (the frequency and sequence of syscalls) over "semantics" (the deep meaning of the code). While semantic models are powerful, they are often too slow for the rapid update cycles of modern ADVs. Multi-IF proves that syntax-based detection, when optimized through multi-feature clusters, provides a practical and "inexpensive" security layer.

Future research will focus on intercepting higher-level APIs, such as those within the Robot Operating System (ROS), to create a more precise model of driving logic. Additionally, the researchers aim to incorporate contextual awareness—mapping syscall sequences to specific vehicle actions like "turning left"—to further reduce false positives in dynamic environments.

Summary updated

Abstract

Autonomous driving vehicles (ADVs) are implemented with rich software functions and equipped with many sensors, which in turn brings broad attack surface. Moreover, the execution environment of ADVs is often open and complex. Hence, ADVs are always at risk of safety and security threats. This paper proposes a fast method called Multi-IF, using multiple invocation features of system calls to detect anomalies in self-driving systems. Since self-driving functions take most of the computation resources and upgrade frequently, Multi-IF is designed to work under such resource constraints and support frequent updates. Given the collected sequences of system calls, the combination of different syntax patterns is used to analyze and construct feature vectors of those sequences. By taking the feature vectors as inputs, one-class support vector machine is adopted to determine whether the current sequence of system calls is abnormal, which is trained with the feature vectors from the normal sequences. The evaluations on both simulated and real data prove that the proposed method is effective in identifying the abnormal behavior after minutes of feature extraction and training. Further comparisons with the existing methods on the ADFA-LD data set also validate that the proposed approach achieves a higher accuracy with less time overhead.
BibTeX
@misc{cheng2020multi,
  author = {Kun Cheng and Yuebin Bai and Yuan Zhou and Chao Yu and Yang Liu},
  title = {{Multi-IF : An Approach to Anomaly Detection in Self-Driving Systems}},
  year = {2020},
  eprint = {2004.12581},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2004.12581},
}
Figure