Anonymizing k-Facial Attributes via Adversarial Perturbations

Saheb Chhabra, Richa Singh, Mayank Vatsa, Gaurav Gupta

Abstract: A face image not only provides details about the identity of a subject but also reveals several attributes such as gender, race, sexual orientation, and age. Advancements in machine learning algorithms and popularity of sharing images on the World Wide Web, including social media websites, have increased the scope of data analytics and information profiling from photo collections. This poses a serious privacy threat for individuals who do not want to be profiled. This research presents a novel algorithm for anonymizing selective attributes which an individual does not want to share without affecting the visual quality of images. Using the proposed algorithm, a user can select single or multiple attributes to be surpassed while preserving identity information and visual content. The proposed adversarial perturbation based algorithm embeds imperceptible noise in an image such that attribute prediction algorithm for the selected attribute yields incorrect classification result, thereby preserving the information according to user's choice. Experiments on three popular databases i.e. MUCT, LFWcrop, and CelebA show that the proposed algorithm not only anonymizes k-attributes, but also preserves image quality and identity information.

1. Introduction

Modern facial recognition systems do more than identify individuals; they extract "soft biometric" attributes such as age, gender, and ethnicity. This capability enables large-scale information profiling, posing a significant privacy risk. Traditional anonymization techniques like blurring or pixelation often degrade image quality to the point of making the photo unusable for legitimate identity verification. This research proposes a method to selectively anonymize $k$ facial attributes using adversarial perturbations, ensuring that chosen traits are hidden from AI classifiers while preserving the image's visual fidelity and the subject's identity for human observers and authorized recognition systems.

2. Methodology

The proposed algorithm utilizes adversarial perturbations—mathematically calculated, imperceptible noise—to mislead attribute classifiers. Unlike generative methods that morph faces, this approach "pushes" the original image's features across the classification boundary of the target attribute without requiring a reference face.

Figure 5: The proposed attribute anonymization framework uses an iterative optimization loop to perturb the input image until the target attribute is misclassified while preserving other features.

The core of the system is an optimization-based attack inspired by the Carlini & Wagner framework. The loss function is designed to achieve three primary objectives:

  • Attribute Suppression: Forcing the classifier to assign an incorrect label to the selected $k$ attributes.
  • Attribute Preservation: Maintaining the correct classification for attributes the user chooses to share.
  • Identity and Visual Fidelity: Minimizing the $L_2$ distance between the original and perturbed images and ensuring identity features remain consistent to support face matching.

3. Experimental Evaluation

The method was rigorously tested on three benchmarks:

  • MUCT: Focused on gender suppression.
  • LFWcrop: Evaluated identity preservation during the anonymization process.
  • CelebA: Tested complex scenarios involving the simultaneous suppression and preservation of multiple attributes among its 40 annotated traits.

4. Results and Discussion

The algorithm demonstrated exceptional performance in both suppression and preservation. On the CelebA dataset, detection accuracy for attributes like "Attractive" and "Smiling" plummeted from high baseline levels to near zero (e.g., 89.31% to 0.28% for "Attractive"). In the MUCT dataset, the system achieved a 100% misclassification rate for target attributes.

Figure 4: Qualitative results showing original images and their anonymized counterparts. The perturbations are virtually invisible to the human eye, maintaining high visual quality and identity across various demographics.

Critically, identity preservation remained high. Evaluation using ROC and CMC curves showed that the Genuine Acceptance Rate for face recognition systems (including black-box models like OpenFace and VGGFace) was minimally affected, allowing the images to remain functional for secure authentication.

5. Conclusion

This research provides a robust framework for selective privacy in digital imagery. By turning adversarial machine learning into a defensive tool, users can regain control over their personal data. While the current optimization process is computationally intensive, future work will focus on improving efficiency and enhancing robustness against common image transformations like JPEG compression and resizing.

Summary updated

Abstract

A face image not only provides details about the identity of a subject but also reveals several attributes such as gender, race, sexual orientation, and age. Advancements in machine learning algorithms and popularity of sharing images on the World Wide Web, including social media websites, have increased the scope of data analytics and information profiling from photo collections. This poses a serious privacy threat for individuals who do not want to be profiled. This research presents a novel algorithm for anonymizing selective attributes which an individual does not want to share without affecting the visual quality of images. Using the proposed algorithm, a user can select single or multiple attributes to be surpassed while preserving identity information and visual content. The proposed adversarial perturbation based algorithm embeds imperceptible noise in an image such that attribute prediction algorithm for the selected attribute yields incorrect classification result, thereby preserving the information according to user's choice. Experiments on three popular databases i.e. MUCT, LFWcrop, and CelebA show that the proposed algorithm not only anonymizes k-attributes, but also preserves image quality and identity information.
BibTeX
@inproceedings{chhabra2018anonymizing,
  author = {Saheb Chhabra and Richa Singh and Mayank Vatsa and Gaurav Gupta},
  title = {{Anonymizing k-Facial Attributes via Adversarial Perturbations}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/91},
}
Figure