SkeletonVis: Interactive Visualization for Understanding Adversarial Attacks on Human Action Recognition Models

Haekyu Park, Zijie J. Wang, Nilaksh Das, Anindya S. Paul, Pruthvi Perumalla, Zhiyan Zhou, Duen Horng Chau

Abstract

Skeleton-based human action recognition technologies are increasingly used in video-based applications, such as home robotics, healthcare for the aging population, and surveillance. However, such models are vulnerable to adversarial attacks, raising serious concerns for their use in safety-critical applications. To develop an effective defense against attacks, it is essential to understand how such attacks mislead pose detection models into making incorrect predictions. We present SkeletonVis, the first interactive system that visualizes how attacks work on these models to enhance human understanding of adversarial vulnerabilities.

Introduction: The Vulnerability of Skeleton-Based Recognition

Human action recognition—the ability to identify actions like "walking," "lunging," or "sitting"—is a critical component in safety-sensitive fields like autonomous driving, home robotics, and eldercare. Many state-of-the-art models utilize "skeletons," which are the 2D or 3D coordinates of human joints, to make these predictions.

Despite their efficiency, these models are highly vulnerable to adversarial attacks. These attacks involve tiny, often imperceptible changes to video pixels that cause the underlying pose detection model to hallucinate incorrect joint positions. This leads to dangerous misclassifications by the action recognition model. Because deep learning models often function as "black boxes," it is difficult for developers to understand why an attack succeeded. SkeletonVis addresses this gap by providing a visual interface to debug and analyze these failures.

System Design and Methods

The SkeletonVis system is built around a standard two-stage action recognition pipeline. First, a pose detection model (Detectron2 Keypoint R-CNN) identifies 17 human joints in every video frame. Second, an action classification model (Spatial Temporal Graph Convolutional Network, or ST-GCN) analyzes the movement of these joints over time to classify the action.

SkeletonVis allows users to compare a "benign" (normal) video with its "attacked" version, typically generated using the Fast Gradient Method (FGM). The tool features three primary interactive views:

  1. Comparison View (A-1): Overlays the benign and attacked skeletons in a single frame to show spatial displacement.
  2. Split View (A-2): Visualizes the "trajectory" of joints over time using transparency to show movement, making it easy to see if a joint's path is smooth or erratic.
  3. Timeline View (B): Provides a quantitative line chart of "Average Joint Displacement," helping users instantly spot "spikes" where the attack is most aggressive.
The SkeletonVis interface provides a comprehensive overview of adversarial impact through Comparison, Split, and Timeline views.

Key Insights: Visualizing the "Invisible" Attack

Through the use of SkeletonVis on the UCF-101 action classification dataset, several key insights into adversarial behavior have emerged:

Spurious Jumping and Temporal Incoherence

One of the most striking findings is that adversarial attacks often cause joints to "jump" spuriously from one frame to the next. While the video might look normal to a human eye, the underlying skeleton data becomes temporally incoherent. This incoherence is precisely what confuses the ST-GCN model, which relies on smooth temporal transitions to recognize actions.

Targeted Vulnerabilities

The tool reveals that attacks do not necessarily perturb the whole body. For instance, in a "lunge" video, an attack might specifically manipulate the joints of the left foot to trick the model into predicting "exercising with a gym ball." This targeted nature suggests that models may have specific "weak points" in their spatial understanding.

Identifying Exploited Frames

By using the Timeline View, researchers can see that attacks are not uniform across a video. There are specific "spike" frames where the displacement is highest; these are the frames the attack "hones in on" to maximize its impact on the final classification.

The Timeline View highlights specific frames where joint displacement spikes, indicating the most aggressive moments of an adversarial attack.

Results and Impact

SkeletonVis successfully demonstrates how a basic adversarial attack can manipulate a complex model.

  • Qualitative Success: The tool makes the "invisible" attack visible. By showing "Skeleton Trajectory Not Coherent," it provides an intuitive explanation for model failure that raw data or accuracy scores cannot provide.
  • Interactive Debugging: The interface allows users to scrub through timelines, pause at specific frames of interest, and toggle between different joint trajectories, effectively turning a complex forensic task into an interactive exploration.

Limitations and Future Work

While SkeletonVis is a powerful diagnostic tool, several areas remain for future exploration. The current demonstration focuses on the ST-GCN architecture; future versions could be adapted to other action recognition models like 3D CNNs. Additionally, understanding how more sophisticated, "stealthy" attacks appear in the visualization would be a valuable next step. Finally, the scalability of the tool to videos with multiple people or highly cluttered backgrounds remains to be addressed.

Summary updated

Abstract

Skeleton-based human action recognition technologies are increasingly used in video-based applications, such as home robotics, healthcare on the aging population, and surveillance. However, such models are vulnerable to adversarial attacks, raising serious concerns for their use in safety-critical applications. To develop an effective defense against attacks, it is essential to understand how such attacks mislead the pose detection models into making incorrect predictions. We present SkeletonVis, the first interactive system that visualizes how the attacks work on the models to enhance human understanding of attacks.
BibTeX
@inproceedings{park2021skeletonvis,
  author = {Haekyu Park and Zijie J. Wang and Nilaksh Das and Anindya S. Paul and Pruthvi Perumalla and Zhiyan Zhou and Duen Horng Chau},
  title = {{SkeletonVis: Interactive Visualization for Understanding Adversarial Attacks on Human Action Recognition Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=x1Yf6s-grYJ},
}
Figure