Proactive Privacy-preserving Learning for Retrieval

Peng-Fei Zhang, Zi Huang, Xin-Shun Xu

Abstract

Proactive Privacy-preserving Learning (PPL) is a novel framework designed to protect image privacy against malicious retrieval systems. By subtly modifying images before they are uploaded, PPL creates quasi-imperceptible perturbations that significantly degrade the performance of unauthorized search engines while maintaining utility for legitimate users. This paper details the adversarial training process between a generator and a surrogate model, demonstrating robust protection across various datasets and neural network architectures.

Introduction

Deep Neural Networks (DNNs) have revolutionized image retrieval, enabling efficient searches across vast online databases. However, this technology also poses significant privacy risks. Malicious actors can use a single query photo to scrape a person's digital footprint or build unauthorized databases without consent. Existing privacy methods often focus on encryption or model protection, but there is a growing need for data-centric solutions that allow users to protect their raw images proactively before they are released online.

Overview of the PPL threat model and proactive defense mechanism.

The PPL Framework

The proposed PPL framework treats privacy protection as a competitive two-player game between an adversarial generator ($\mathcal{G}$) and a surrogate retrieval model ($\mathcal{F}$).

Detailed architecture of the PPL framework showing the adversarial generator and surrogate model.

The framework consists of several key components:

  • Adversarial Generator ($\mathcal{G}$): An auto-encoder-like model that adds a tiny, invisible noise pattern to a clean image. It aims to maximize a Privacy-preserving Loss ($J_p$) to confuse retrieval models.
  • Surrogate Retrieval Model ($\mathcal{F}$): Since the target malicious model is unknown, PPL trains against a surrogate that attempts to remain robust against the generator's perturbations.
  • Adversarial Training: The generator and surrogate are trained simultaneously. A Gradient Reversal Layer (GRL) is employed to flip gradients during backpropagation, facilitating the opposing optimization objectives in a single step.
  • Affinity-preserving Loss ($J_o$): This loss ensures that the legitimate similarity structure of the data is preserved, allowing authorized users to maintain high retrieval accuracy even with perturbed images.

Experimental Evaluation

The researchers evaluated PPL on three widely recognized benchmarks:

  • CIFAR-10: A collection of 60,000 small images across 10 classes.
  • ImageNet: A subset of 13 categories (approx. 18,000 images) from the industry-standard dataset.
  • FashionMNIST: 70,000 grayscale images of clothing items.

Results and Key Insights

The effectiveness was measured using mean Average Precision (mAP). When tested against malicious models like ADSH or DSDH:

  • Privacy Protection: Retrieval accuracy plummeted by an average of 50.8% to 70.7% on CIFAR-10 and 43.4% to 62.9% on FashionMNIST.
  • Visual Quality: The perturbations were constrained by an $l_\infty$ norm, meaning the changes are virtually invisible to the human eye.
  • Transferability: A major finding is that perturbations designed to fool the surrogate model often successfully fool other architectures (e.g., ResNet18, ResNet50, WideResNet, MobileNet, GoogLeNet), proving the method's real-world robustness.
  • Utility vs. Privacy: By carefully balancing loss functions, the adversarial images remain useful for authorized "clean-to-clean" or "adversarial-to-adversarial" searches.

Conclusion

PPL demonstrates that adversarial examples can be repurposed as a proactive defense mechanism for user privacy. While highly effective, future work is needed to improve performance on high-resolution datasets like ImageNet, reduce computational overhead for end-users, and address potential adaptive attacks where malicious actors might develop specific "anti-PPL" techniques.

Summary updated

Abstract

Deep Neural Networks (DNNs) have recently achieved remarkable performance in image retrieval, yet posing great threats to data privacy. On the one hand, one may misuse a deployed DNNs based system to look up data without consent. On the other hand, organizations or individuals would legally or illegally collect data to train high-performance models outside the scope of legitimate purposes. Unfortunately, less effort has been made to safeguard data privacy against malicious uses of DNNs. In this paper, we propose a data-centric Proactive Privacy-preserving Learning (PPL) algorithm for hashing based retrieval, which achieves the protection purpose by employing a generator to transfer the original data into the adversarial data with quasi-imperceptible perturbations before releasing them. When the data source is infiltrated, the adversarial data can confuse menacing retrieval models to make erroneous predictions. Given that the prior knowledge of malicious models is not available, a surrogate retrieval model is instead introduced acting as a fooling target. The framework is trained by a two-player game conducted between the generator and the surrogate model. More specifically, the generator is updated to enlarge the gap between the adversarial data and the original data, aiming to lower the search accuracy of the surrogate model. On the contrary, the surrogate model is trained with the opposing objective that is to maintain the search performance. As a result, an effective and robust adversarial generator is encouraged. Furthermore, to facilitate an effective optimization, a Gradient Reversal Layer (GRL) module is inserted to connect two models, enabling the two-player game in a one-step learning. Extensive experiments on three widely-used realistic datasets prove the effectiveness of the proposed method.
BibTeX
@inproceedings{zhang2021proactive,
  author = {Peng-Fei Zhang and Zi Huang and Xin-Shun Xu},
  title = {{Proactive Privacy-preserving Learning for Retrieval}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i4.16449},
}
Figure