Self-recoverable Adversarial Examples: A New Effective Protection Mechanism in Social Networks

Jiawei Zhang, Jinwei Wang, Hao Wang, Xiangyang Luo

Abstract

Malicious intelligent algorithms greatly threaten the security of social users' privacy by detecting and analyzing the uploaded photos to social network platforms. The destruction to DNNs brought by the adversarial attack sparks the potential that adversarial examples serve as a new protection mechanism for privacy security in social networks. However, the existing adversarial example does not have recoverability for serving as an effective protection mechanism. To address this issue, we propose a recoverable generative adversarial network to generate self-recoverable adversarial examples. By modeling the adversarial attack and recovery as a united task, our method can minimize the error of the recovered examples while maximizing the attack ability, resulting in better recoverability of adversarial examples. To further boost the recoverability of these examples, we exploit a dimension reducer to optimize the distribution of adversarial perturbation. The experimental results prove that the adversarial examples generated by the proposed method present superior recoverability, attack ability, and robustness on different datasets and network architectures, which ensure its effectiveness as a protection mechanism in social networks.

Introduction

In the age of social media, every photo uploaded is a potential goldmine for malicious intelligent algorithms. Deep Neural Networks (DNNs) can automatically detect locations, identify individuals, and catalog property without user consent. While adversarial examples—images tweaked with imperceptible noise to fool AI—have existed for years, they have traditionally been a one-way street. Once noise is added to reliably break a classifier, the image is often degraded for the user as well. This research addresses the challenge of recoverability: creating a "digital mask" aggressive enough to stop AI but removable near-losslessly by the rightful owner.

Methodology: The Recoverable GAN (RGAN)

The core contribution is the Recoverable Generative Adversarial Network (RGAN). Unlike traditional methods that treat attacking a model and recovering an image as separate problems, RGAN models them as a united task.

The architecture consists of three main components:

  1. The Generator ($G$): An encoder-decoder setup that creates the adversarial perturbation.
  2. The Dimension Reducer ($DR$): This component uses down-sampling and up-sampling with skip connections to simplify the noise. The researchers discovered that "structured" noise—even if intense—is much easier for a recovery network to handle than "messy" noise.
  3. The Recoverer ($R$): A network mirrored after the generator’s architecture. Trained jointly with the generator, it learns to subtract the specific "flavor" of noise added by $G$.
The self-recoverable adversarial example pipeline, showing the original image, generated perturbations, the adversarial example, and the final recovered image.

By optimizing for both attack ability and recoverability simultaneously, the system finds a balance that previous one-off adversarial attacks could not achieve.

Experimental Setup

The approach was evaluated using two primary benchmarks:

  • MNIST: Handwritten digits (28x28 pixels) for basic functional testing.
  • Caltech-256: Complex real-world imagery (224x224 pixels) across 256 object categories.

The team tested the "transferability" of their protection across several popular architectures, including LeNet-5, ResNet-50, DenseNet-121, and MobileNetV3.

Key Insights and Results

Complexity vs. Intensity

A significant finding is that the complexity of adversarial noise matters more for recovery than its intensity. By using the Dimension Reducer to create "coarser" noise, the authors maintained high attack success rates while making the recovery process significantly more accurate.

Homologous Recovery

The recovery network performs best when it mirrors the generator's structure (homology). This allows the recoverer to better extract and nullify the specific features added during the attack phase.

Performance Metrics

The RGAN outperformed existing state-of-the-art combinations:

  • Superior Recovery: On Caltech-256 (ResNet-50), the RGAN achieved a PSNR (Peak Signal-to-Noise Ratio) of 43.73 dB, compared to 30-34 dB for competitive methods. The recovered image is virtually indistinguishable from the original.
  • High Attack Success: The system maintained an Attack Success Rate (ASR) of 90-95% against undefended models.
  • Robustness: The protection remains effective after common social media processing like JPEG compression, Gaussian blurring, or resizing, misleading target AIs 71% to 94% of the time.

Conclusion

While RGAN provides a powerful new privacy tool, future work includes scaling the model to deeper architectures and optimizing computational overhead for real-time social network deployment. The current trade-off between robustness and raw attack success also presents an avenue for further refinement.

Summary updated

Abstract

Malicious intelligent algorithms greatly threaten the security of social users' privacy by detecting and analyzing the uploaded photos to social network platforms. The destruction to DNNs brought by the adversarial attack sparks the potential that adversarial examples serve as a new protection mechanism for privacy security in social networks. However, the existing adversarial example does not have recoverability for serving as an effective protection mechanism. To address this issue, we propose a recoverable generative adversarial network to generate self-recoverable adversarial examples. By modeling the adversarial attack and recovery as a united task, our method can minimize the error of the recovered examples while maximizing the attack ability, resulting in better recoverability of adversarial examples. To further boost the recoverability of these examples, we exploit a dimension reducer to optimize the distribution of adversarial perturbation. The experimental results prove that the adversarial examples generated by the proposed method present superior recoverability, attack ability, and robustness on different datasets and network architectures, which ensure its effectiveness as a protection mechanism in social networks.
BibTeX
@article{zhang2023self,
  author = {Jiawei Zhang and Jinwei Wang and Hao Wang and Xiangyang Luo},
  title = {{Self-recoverable Adversarial Examples: A New Effective Protection Mechanism in Social Networks}},
  journal = {IEEE Trans. Circuits Syst. Video Technol.},
  year = {2023},
  url = {https://openreview.net/forum?id=q259DuI9qgJ},
}
Figure