Federated Learning without Revealing the Decision Boundaries

Roozbeh YousefzadeharXiv 2021 · arXiv

Abstract

In traditional Federated Learning (FL), privacy techniques often focus on obscuring raw data while inadvertently exposing the model's decision boundaries to a central server. This exposure creates a significant security vulnerability, allowing potential adversaries to map the model's internal logic and launch targeted attacks. This paper proposes a novel theoretical framework that utilizes shearlet-based encryption and an integrated decryption module within the neural network architecture. By training on encrypted shearlet coefficients, the proposed method aims to preserve "boundary privacy," ensuring that neither the raw data nor the model's functional decision-making process is revealed to the coordinating entity.

1. Introduction

Federated Learning (FL) enables multiple participants to collaboratively train a machine learning model without sharing their local datasets. While this decentralization offers a layer of privacy, the central server still coordinates the aggregation of model updates. To prevent the server from reconstructing private data from these updates, researchers have developed various "mixing" or "masking" techniques. A prominent example is the InstaHide method, which blends multiple images into unrecognizable noise before transmission.

However, a critical security gap remains. While these methods may hide the "treasure" (the raw image content), they often provide the server with a "map" (the model's decision boundaries). This paper addresses this vulnerability by introducing a framework designed to protect the functional logic of the model itself.

2. The Problem: Boundary Leakage and the "Flip Label" Trap

The fundamental flaw in current image-mixing privacy methods lies in the nature of the training samples they generate. When images are mixed or interpolated—often referred to as using "flip labels"—the resulting data points are, by definition, located at the transition points between classes.

In mathematical terms, these points reside on or near the model's decision boundaries. If a central server receives a large volume of these boundary-adjacent samples, it can effectively reconstruct the geometry of the model's decision space. Once the server understands where the model "decides" between different classifications, it can launch sophisticated adversarial attacks, manipulating inputs to cross these boundaries with minimal perturbation. Thus, protecting the data's appearance is insufficient if the model's internal boundaries remain transparent.

3. Proposed Methodology: Shearlet-Based Encryption

To mitigate boundary leakage, we propose a shift from simple image mixing to a robust encryption-decryption architecture based on shearlet systems.

3.1 Shearlet Systems for Feature Representation

Shearlets are a multi-scale mathematical framework that extends the capabilities of traditional wavelets. While wavelets are effective at identifying point-wise singularities, shearlets are specifically designed to handle "anisotropic" features. This makes them exceptionally well-suited for representing edges and complex geometric shapes within images—the very features that neural networks rely on for classification.

3.2 Encryption at the Source

In the proposed workflow, image owners do not send raw or mixed pixels. Instead, they use a shearlet system to decompose their images into a set of complex-valued coefficients. This decomposition acts as a form of encryption. Without the specific "key"—defined by the unique shearlet function $\psi$ and the specific filter boundaries used—reversing these coefficients back into the original pixel space is computationally prohibitive.

3.3 The Hidden Decryption Module

The neural network architecture is modified to include a "hidden" decryption layer at its entry point. This layer is initialized with the secret shearlet key, allowing it to internally reconstruct the necessary features for training.

  1. Transmission: The server receives only the encrypted shearlet coefficients.
  2. Internal Processing: The model performs an internal decryption before the standard convolutional layers.
  3. Learning: The learning process occurs in a way that the server observes updates to the weights, but because the input space is encrypted, the server cannot map these updates back to specific locations in the original pixel-based decision space.

4. Key Insights

The proposed framework rests on the distinction between data privacy and functional security. By using complex-valued coefficients and truncated filter boundaries, the "search space" for a potential attacker attempting to decrypt the data becomes massive. Furthermore, by avoiding the use of explicit "flip labels" in the raw input space, the method prevents the server from identifying the exact location of the decision boundaries.

5. Limitations and Future Work

As a preliminary theoretical framework, several challenges remain:

  • Key Management: The security of the system is entirely dependent on the secrecy of the shearlet function $\psi$. Future research must determine optimal methods for generating and distributing these keys.
  • Cryptographic Rigor: While the inversion of shearlet coefficients is theoretically difficult, a formal cryptographic proof of the security margin is still required.
  • Computational Complexity: The addition of an internal decryption module increases the computational overhead of the model. Strategies for optimizing this layer, or potentially removing it during the inference phase, need to be explored.

6. Conclusion

Protecting the privacy of participants in Federated Learning requires more than just obscuring raw data; it requires securing the model's decision boundaries. By leveraging the mathematical properties of shearlet systems, this proposed framework offers a path toward a more secure FL environment where the central server remains blind to both the data and the "map" of the model's internal logic.

Summary updated

Abstract

We consider the recent privacy preserving methods that train the models not on original images, but on mixed images that look like noise and hard to trace back to the original images. We explain that those mixed images will be samples on the decision boundaries of the trained model, and although such methods successfully hide the contents of images from the entity in charge of federated learning, they provide crucial information to that entity about the decision boundaries of the trained model. Once the entity has exact samples on the decision boundaries of the model, they may use it for effective adversarial attacks on the model during training and/or afterwards. If we have to hide our images from that entity, how can we trust them with the decision boundaries of our model? As a remedy, we propose a method to encrypt the images, and have a decryption module hidden inside the model. The entity in charge of federated learning will only have access to a set of complex-valued coefficients, but the model will first decrypt the images and then put them through the convolutional layers. This way, the entity will not see the training images and they will not know the location of the decision boundaries of the model.
BibTeX
@misc{yousefzadeh2021federated,
  author = {Roozbeh Yousefzadeh},
  title = {{Federated Learning without Revealing the Decision Boundaries}},
  year = {2021},
  eprint = {2103.00695},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.00695},
}
Figure