A Training-based Identification Approach to VIN Adversarial Examples

Yingdi Wang, Wenjia Niu, Tong Chen, Yingxiao Xiang, Jingjing Liu, Gang Li, Jiqiang LiuarXiv 2018 · arXiv

Abstract

Adversarial attacks on Deep Reinforcement Learning (DRL) models, particularly Value Iteration Networks (VIN) used in robot path planning, present significant security risks. Identifying successful attacks currently relies on slow, manual human inspection. This paper introduces a hybrid automated pipeline that combines geometric logic with Support Vector Machine (SVM) image classification to categorize attack outcomes. The system achieves 97% accuracy and processes data over 1,000 times faster than human observers, providing a scalable solution for adversarial detection in autonomous navigation.

Introduction

In AI security, adversarial examples are well-documented in image recognition. However, their impact on robotics and path planning is a burgeoning field of concern. Value Iteration Networks (VIN) are popular DRL models that allow robots to learn navigation policies. Attackers can exploit these models by adding "invisible" obstacles—minor map tweaks that appear benign to humans but cause the robot to fail or deviate significantly from an optimal path.

The Adversarial Challenge in Path Planning

The primary difficulty in studying these attacks is determining their success. Unlike simple classification tasks, path planning is continuous. A successful attack might cause a robot to never reach its goal or to take a vastly different route. Historically, researchers have manually compared map pairs to evaluate attack efficacy, a process that is subjective and unscalable.

Comparison between a normal map (a) and an adversarial map (b) featuring a single strategic obstacle.

Proposed Methodology

To automate this detection, we propose a two-stage hybrid pipeline. The system classifies adversarial results into four categories: Unreached Path (UrP), Fork Path (FP), Detour Path (DP), and Unchanged Path (UcP).

Stage 1: Logic-based Filtering

The first stage employs simple coordinate-based logic. If the robot's final position does not match the goal, it is classified as an UrP (successful attack). If the entire path sequence is identical to the original, it is a UcP (unsuccessful attack). This stage handles "easy" cases with minimal computational cost.

Stage 2: SVM-based Image Classification

For cases where the path has changed but still reaches the goal, the system must distinguish between a "Fork Path" (a major deviation) and a "Detour Path" (a minor, negligible adjustment).

Visual comparison of an original path (a) and an adversarial path (b) demonstrating a significant deviation or 'fork'.

The paths are converted into 28x28 grayscale images. Interestingly, a traditional Support Vector Machine (SVM) was found to outperform Convolutional Neural Networks (CNNs) for this task. The sparse nature of path images—where most pixels are empty—favors the SVM's ability to find clear decision boundaries in lower-dimensional feature spaces, whereas CNNs struggle to extract meaningful features from such sparse data.

Experimental Results

The system was evaluated on a synthetic dataset of 5,000 map pairs. To address class imbalance (as successful "Fork Paths" were less frequent), data augmentation techniques like flipping and rotation were applied.

  • Accuracy: The SVM-based classifier achieved a test accuracy of 97%.
  • Efficiency: The automated pipeline processed 100 path pairs in 0.22 seconds, compared to approximately 300 seconds for a human observer.
  • Robustness: Precision-Recall and ROC curves confirmed the model's high confidence and reliability across different attack scenarios.

Conclusion and Future Work

This hybrid approach demonstrates that combining simple geometric logic with classical machine learning can effectively solve complex detection problems in DRL security. While currently limited to 28x28 grid maps, future research will focus on scaling this threshold-based detection to high-resolution maps and continuous 3D environments. The framework's success with VIN suggests potential applicability to other DRL algorithms such as DQN or A3C.

Summary updated

Abstract

With the rapid development of Artificial Intelligence (AI), the problem of AI security has gradually emerged. Most existing machine learning algorithms may be attacked by adversarial examples. An adversarial example is a slightly modified input sample that can lead to a false result of machine learning algorithms. The adversarial examples pose a potential security threat for many AI application areas, especially in the domain of robot path planning. In this field, the adversarial examples obstruct the algorithm by adding obstacles to the normal maps, resulting in multiple effects on the predicted path. However, there is no suitable approach to automatically identify them. To our knowledge, all previous work uses manual observation method to estimate the attack results of adversarial maps, which is time-consuming. Aiming at the existing problem, this paper explores a method to automatically identify the adversarial examples in Value Iteration Networks (VIN), which has a strong generalization ability. We analyze the possible scenarios caused by the adversarial maps. We propose a training-based identification approach to VIN adversarial examples by combing the path feature comparison and path image classification. We evaluate our method using the adversarial maps dataset, show that our method can achieve a high-accuracy and faster identification than manual observation method.
BibTeX
@misc{wang2018training,
  author = {Yingdi Wang and Wenjia Niu and Tong Chen and Yingxiao Xiang and Jingjing Liu and Gang Li and Jiqiang Liu},
  title = {{A Training-based Identification Approach to VIN Adversarial Examples}},
  year = {2018},
  eprint = {1810.08070},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1810.08070},
}
Figure