Generative Gradient Inversion via Over-Parameterized Networks in Federated Learning

Chi Zhang, Xiaoman Zhang, Ekanut Sotthiwat, Yanyu Xu, Ping Liu, Liangli Zhen, Yong LiuICCV 2023 · DOI

Abstract

Federated Learning (FL) is widely regarded as a privacy-preserving framework because it keeps raw data on local devices, sharing only model gradients. However, recent "gradient inversion" attacks have challenged this assumption. This paper presents CI-Net, a novel attack strategy that leverages over-parameterized generative networks to reconstruct private training data with high fidelity. Unlike previous methods, CI-Net succeeds at large batch sizes and high resolutions without requiring hand-crafted image priors. Our results on CIFAR-10 and ImageNet demonstrate that CI-Net significantly outperforms state-of-the-art methods, highlighting critical vulnerabilities in current FL privacy guarantees.

Introduction

The core promise of Federated Learning is that "data stays local." By only communicating model updates (gradients), the central server theoretically never sees the user's private images. Gradient inversion attacks attempt to break this by finding images that produce the observed gradients. Historically, these attacks struggled with two main issues: scaling to large batch sizes or high resolutions, and a heavy dependence on manual "image priors" (like Total Variation) which often led to noisy or failed reconstructions if misconfigured.

Methods: The CI-Net Architecture

The authors propose the Convolutional Inversion Network (CI-Net). Rather than optimizing image pixels directly, the attack optimizes the parameters of a generative network. This network is designed to output images that, when processed by the target model, generate gradients matching those intercepted from the user.

The CI-Net architecture utilizes a progressive-growing structure and a gradient matching loop to reconstruct images.

Key technical features include:

  • Over-Parameterization: The generative model is intentionally designed with more parameters than the target model, ensuring a sufficiently large search space to find valid image matches.
  • Implicit Convolutional Regularization: By using a CNN structure, the model naturally favors "natural-looking" images over noise, effectively acting as an implicit prior (similar to Deep Image Prior).
  • Progressive-Growing Architecture: The network starts with low-resolution cores and progressively scales up using nearest-neighbor interpolation, which preserves pixel-level details better than linear methods.

Experimental Results

CI-Net was evaluated against several baseline methods, including DLG, iDLG, IGT, and GIAS, using the CIFAR-10 and ImageNet datasets.

Quantitative SSIM comparison across batch sizes (BS) and qualitative visual reconstruction results.

As shown in the results, CI-Net maintains high reconstruction quality even as the batch size increases. On CIFAR-10 with a batch size of 128, CI-Net achieved a mean PSNR of 31.40, whereas the best competitor (GIAS) only reached ~11.05. On high-resolution ImageNet data (256x256) with a batch size of 24, CI-Net achieved an SSIM of 0.52, while all other methods failed (SSIM < 0.05). The stability of the attack across different random seeds further confirms its robustness.

Discussion and Conclusion

The success of CI-Net reveals that gradient matching is not a unique problem; many images can produce the same gradient. CI-Net's architectural bias toward natural images allows it to find the "correct" solution where others fail. Furthermore, the study suggests that traditional regularization terms can sometimes be counterproductive, shifting the optimization away from the true data.

While CI-Net is computationally more expensive (requiring ~40% more GPU memory) and can produce slightly blurred results at very high resolutions, it represents a significant leap in the feasibility of gradient inversion attacks. These findings underscore the need for more robust defense mechanisms in Federated Learning, as simple gradient sharing is clearly insufficient for data privacy.

Summary updated

BibTeX
@inproceedings{zhang2023generative,
  author = {Chi Zhang and Xiaoman Zhang and Ekanut Sotthiwat and Yanyu Xu and Ping Liu and Liangli Zhen and Yong Liu},
  title = {{Generative Gradient Inversion via Over-Parameterized Networks in Federated Learning}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
  doi = {10.1109/iccv51070.2023.00473},
}
Figure