Shape Evasion: Preventing Body Shape Inference of Multi-Stage Approaches

Hosnieh Sattar, Katharina Krombholz, Gerard Pons-Moll, Mario Fritz

Abstract

Modern approaches to pose and body shape estimation have recently achieved strong performance even under challenging real-world conditions. Even from a single image of a clothed person, a realistic looking body shape can be inferred that captures a users' weight group and body shape type well. This opens up a whole spectrum of applications -- in particular in fashion -- where virtual try-on and recommendation systems can make use of these new and automatized cues. However, a realistic depiction of the undressed body is regarded highly private and therefore might not be consented by most people. Hence, we ask if the automatic extraction of such information can be effectively evaded. While adversarial perturbations have been shown to be effective for manipulating the output of machine learning models -- in particular, end-to-end deep learning approaches -- state of the art shape estimation methods are composed of multiple stages. We perform the first investigation of different strategies that can be used to effectively manipulate the automatic shape estimation while preserving the overall appearance of the original image.

Introduction

Recent breakthroughs in computer vision have enabled the extraction of highly accurate 3D body meshes (such as the SMPL model) from standard 2D photographs. While these technologies power innovative applications like virtual fitting rooms, they also present a significant privacy risk. These algorithms can essentially "see through" clothing to estimate a person's undressed body shape, weight, and BMI. This paper explores the feasibility of "Shape Evasion"—using adversarial techniques to prevent these models from accurately inferring private body metrics while maintaining the visual integrity of the original image.

The Privacy Awareness Gap

To understand the social context of this technology, the researchers conducted a user study (N=90). The results revealed a stark "privacy awareness gap." While the vast majority of participants (85-98%) were aware that AI could determine their age or gender from a photo, significantly fewer (47-66%) realized that body mass index (BMI) or detailed 3D shape could be extracted. When informed of this capability, most users reported feeling "extremely uncomfortable" with their 3D shape data being shared, highlighting a critical need for privacy-preserving interventions.

Evading Multi-Stage Shape Estimation

State-of-the-art shape estimation models typically operate in multiple stages: first, a deep learning model (like OpenPose) detects 2D skeletal keypoints; then, an optimization process fits a 3D body model to these points. This multi-stage pipeline is not end-to-end differentiable, making standard adversarial attacks—which rely on computing gradients across the entire network—ineffective.

Overview of the shape evasion pipeline: an intervention on 2D keypoints leads to a generic 3D shape output.

The authors propose targeting the intermediate 2D keypoint detection stage. By subtly manipulating the detected skeleton, the subsequent 3D fitting process can be derailed. They introduce a novel technique called the Masked Fast Gradient Sign Method (MFGSM). Unlike global attacks that apply noise to the entire image, MFGSM applies adversarial perturbations only to the local pixel regions surrounding specific target keypoints. This localized approach is significantly more stealthy, requiring less than half the noise magnitude (0.015 MSE vs. 0.035 MSE) of global attacks to achieve the same level of evasion.

Key Insights and Results

The research identifies that not all joints are equally important for shape estimation. Attacking structural anchors like the hips or shoulders is far more effective than targeting extremities like wrists or elbows. For instance, manually removing hip keypoints in synthetic tests increased shape estimation error by over 60%.

Using the proposed MFGSM attack to suppress keypoints like the knee or elbow successfully increased 3D shape estimation error by approximately 20-21%. Crucially, these attacks do not just "break" the model; they often force the 3D optimizer to default to a generic, average body shape that does not reflect the actual user's physical characteristics, effectively masking their true identity.

Conclusion

This work demonstrates that it is possible to protect body shape privacy through targeted, localized adversarial attacks. By exploiting the vulnerabilities of multi-stage pipelines, the MFGSM method provides a path toward privacy-preserving image sharing. Future work will look into targeted evasion—allowing users to trick models into predicting a specific, false body shape—and investigating the robustness of these attacks against models trained on adversarial data.

Summary updated

Abstract

Modern approaches to pose and body shape estimation have recently achieved strong performance even under challenging real-world conditions. Even from a single image of a clothed person, a realistic looking body shape can be inferred that captures a users' weight group and body shape type well. This opens up a whole spectrum of applications -- in particular in fashion -- where virtual try-on and recommendation systems can make use of these new and automatized cues. However, a realistic depiction of the undressed body is regarded highly private and therefore might not be consented by most people. Hence, we ask if the automatic extraction of such information can be effectively evaded. While adversarial perturbations have been shown to be effective for manipulating the output of machine learning models -- in particular, end-to-end deep learning approaches -- state of the art shape estimation methods are composed of multiple stages. We perform the first investigation of different strategies that can be used to effectively manipulate the automatic shape estimation while preserving the overall appearance of the original image.
BibTeX
@inproceedings{sattar2020shape,
  author = {Hosnieh Sattar and Katharina Krombholz and Gerard Pons-Moll and Mario Fritz},
  title = {{Shape Evasion: Preventing Body Shape Inference of Multi-Stage Approaches}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2020},
  doi = {10.1007/978-3-030-68238-5\_31},
}
Figure