Abstract
While dropout is a staple for regularizing neural networks, recent claims suggested it could also serve as a "free" defense against privacy-shattering gradient inversion attacks in Federated Learning. This paper debunks that hope by introducing the Dropout Inversion Attack (DIA), a method that jointly reconstructs private training data and the random dropout masks used during training. Our results prove that architectural stochasticity alone cannot guarantee data privacy, as DIA achieves high-fidelity reconstructions even when standard attacks fail.
Introduction: The Illusion of Privacy in Federated Learning
In Federated Learning (FL), clients keep their data local and only share model gradients with a central server. For a long time, this was thought to be inherently private. However, Gradient Inversion (GI) attacks have shown that an honest-but-curious server can often "reverse-engineer" the original training images just by looking at those shared gradients.
To mitigate this without destroying model accuracy, researchers have explored Dropout as a defense. By randomly "turning off" neurons during training, dropout makes the model stochastic. The gradients shared by the client are based on a "random version" of the network that the server hasn't seen. Previous attacks failed against dropout-protected models because they couldn't account for this randomness, leading to the belief that dropout might be a viable privacy tool. This paper challenges that assumption: Is the data actually protected, or are existing attacks simply not sophisticated enough?
The Dropout Inversion Attack (DIA)
The authors first demonstrate a "Well-Informed" attack (WIIG), showing that if an attacker knew which neurons were dropped, they could reconstruct the data perfectly. Since a real attacker doesn't know the masks, the authors propose the Dropout Inversion Attack (DIA).
Key components of DIA include:
- Joint Optimization: Unlike standard attacks that only try to find the "dummy image" ($x'$) that produces the observed gradient, DIA treats the dropout masks ($\Psi_A$) as learnable parameters. It attempts to find the image and the mask simultaneously.
- Fuzzy Masks: Binary masks (0 or 1) are non-differentiable and hard to optimize. DIA uses "fuzzy" masks—continuous values between 0 and 1—during the optimization process to allow for gradient-based updates.
- Mask Regularization: To prevent the attack from converging on unrealistic solutions, a regularization term is added to force the average of the learned masks to match the known dropout rate ($p$) of the model.
Experimental Evaluation
The authors evaluated DIA across standard computer vision benchmarks, including MNIST, CIFAR-10, and ImageNet, using architectures ranging from MLPs and CNNs to Vision Transformers (ViT-B/16).
Key Results
The core finding is that stochasticity does not equal privacy. The information is still present in the gradient; it is merely "locked" behind a random key (the mask).
On Vision Transformers (ViT) where standard attacks like "Inverting Gradients" (IG) failed completely (producing unrecognizable noise), DIA achieved high-fidelity reconstructions. For example, on ImageNet with a dropout rate of 0.1, DIA reached an SSIM of 0.72, whereas the standard attack failed. Even with high dropout rates (e.g., $p=0.75$), DIA could still reconstruct images with high structural similarity, though very high rates eventually degrade both reconstruction quality and model utility.
Discussion and Conclusion
The "fuzzy" mask breakthrough allows standard backpropagation to effectively "guess" the client's mask. Interestingly, the authors found that in some cases, dropout might actually make reconstruction easier for a well-informed attacker by simplifying the optimization landscape.
The study concludes that architectural "tricks" like dropout are not a substitute for mathematically rigorous defenses. While DIA is computationally more expensive than standard GI, its success across various architectures highlights a critical vulnerability. For meaningful protection in Federated Learning, dropout must be combined with Differential Privacy (DP) or Secure Aggregation. Future work will explore DIA's performance in other domains, such as Natural Language Processing (NLP).
Summary updated