Protecting Sensitive Data through Federated Co-Training

Amr Abourayya, Jens Kleesiek, Kanishka Rao, Erman Ayday, Bharat Rao, Geoff Webb, Michael Kamp

Abstract

In many critical applications, sensitive data is inherently distributed. Federated learning trains a model collaboratively by aggregating the parameters of locally trained models. This avoids exposing sensitive local data. It is possible, though, to infer upon the sensitive data from the shared model parameters. At the same time, many types of machine learning models do not lend themselves to parameter aggregation, such as decision trees, or rule ensembles. It has been observed that in many applications, in particular healthcare, large unlabeled datasets are publicly available. They can be used to exchange information between clients by distributed distillation, i.e., co-regularizing local training via the discrepancy between the soft predictions of each local client on the unlabeled dataset. This, however, still discloses private information and restricts the types of models to those trainable via gradient-based methods. We propose to go one step further and use a form of federated co-training, where local hard labels on the public unlabeled datasets are shared and aggregated into a consensus label. This consensus label can be used for local training by any supervised machine learning model. We show that this federated co-training approach achieves a model quality comparable to both federated learning and distributed distillation on a set of benchmark datasets and real-world medical datasets. It improves privacy over both approaches, protecting against common membership inference attacks to the highest degree. Furthermore, we show that federated co-training can collaboratively train interpretable models, such as decision trees and rule ensembles, achieving a model quality comparable to centralized training.

Introduction

Federated Learning (FL) was designed to train machine learning models on distributed data without moving that data to a central server. However, the information shared during this process—typically model weights or gradients—can be reverse-engineered. Attackers can perform membership inference attacks to determine if a specific individual's data was used in training, or even reconstruct parts of the private data itself.

Existing semi-supervised FL approaches attempt to mitigate this by sharing "soft labels" (probability distributions) on a public unlabeled dataset. While better than sharing raw weights, soft labels still leak significant signal. This paper introduces FEDCT (Federated Co-Training), a method that shares only definitive "hard labels" to achieve near-optimal privacy while maintaining high accuracy and enabling the use of non-gradient models like Random Forests.

Methodology: Federated Co-Training

The core innovation of FEDCT is a shift from sharing complex model internals to sharing simple, definitive predictions.

  1. The Protocol: Instead of sending model weights, each client uses its local model to predict hard labels (e.g., "This image is a Dog," rather than a probability distribution) for a shared, unlabeled public dataset.
  2. Consensus via Majority Vote: The server collects these hard labels and forms a consensus using a majority vote. These consensus labels then become "pseudo-labels" for the public dataset.
  3. Local Training: Clients download these pseudo-labels and use the newly labeled public data to augment their local private training sets for the next round.
  4. Differential Privacy (DP-FEDCT): To provide formal mathematical privacy guarantees, an XOR-mechanism adds noise to the binary label matrices before they leave the client.
  5. Qualified Majority (QM): To handle data heterogeneity, a quorum requirement (e.g., 90% agreement) ensures only high-confidence pseudo-labels are utilized.

Privacy and Utility Trade-offs

A primary goal of FEDCT is to reduce the empirical privacy vulnerability (VUL) to near 0.5—the theoretical limit equivalent to random guessing. By stripping away the probability distribution (the "confidence" of the model), FEDCT removes the subtle signals that attackers use to reconstruct data.

Privacy-Utility Trade-off: FedCT achieves high accuracy while maintaining vulnerability (VUL) near the theoretical minimum of 0.5.

Experimental Results

The researchers evaluated FEDCT across diverse benchmarks, including image classification (CIFAR10, SVHN), medical imaging (MRI, Pneumonia), and tabular data.

Performance Across Datasets

FEDCT achieved a vulnerability score of ~0.51-0.53 across most datasets, significantly outperforming FedAvg (~0.72) and soft-label sharing (~0.60). Despite the strict privacy, FEDCT matched or exceeded the accuracy of FedAvg on almost all image and medical datasets.

Performance Comparison: FedCT maintains competitive accuracy (ACC) while significantly reducing privacy vulnerability (VUL) across various datasets compared to standard Federated Learning (FedAvg).

Model Agnosticism and Efficiency

Because FEDCT only exchanges labels, it is model-agnostic. Clients can use different architectures—one might use a Random Forest while another uses a Deep Neural Network—and still collaborate. Furthermore, FEDCT is highly communication-efficient. On FashionMNIST, it reduced communication traffic by a factor of approximately 214x compared to FedAvg.

Conclusion

FEDCT demonstrates that sharing "less" information (hard labels instead of weights or soft labels) can actually lead to "more" robust privacy without sacrificing model utility. While it requires access to a relevant unlabeled public dataset, its ability to support heterogeneous, non-gradient models and its superior privacy profile make it a compelling alternative for sensitive distributed data applications.

Summary updated

Abstract

In many critical applications, sensitive data is inherently distributed. Federated learning trains a model collaboratively by aggregating the parameters of locally trained models. This avoids exposing sensitive local data. It is possible, though, to infer upon the sensitive data from the shared model parameters. At the same time, many types of machine learning models do not lend themselves to parameter aggregation, such as decision trees, or rule ensembles. It has been observed that in many applications, in particular healthcare, large unlabeled datasets are publicly available. They can be used to exchange information between clients by distributed distillation, i.e., co-regularizing local training via the discrepancy between the soft predictions of each local client on the unlabeled dataset. This, however, still discloses private information and restricts the types of models to those trainable via gradient-based methods. We propose to go one step further and use a form of federated co-training, where local hard labels on the public unlabeled datasets are shared and aggregated into a consensus label. This consensus label can be used for local training by any supervised machine learning model. We show that this federated co-training approach achieves a model quality comparable to both federated learning and distributed distillation on a set of benchmark datasets and real-world medical datasets. It improves privacy over both approaches, protecting against common membership inference attacks to the highest degree. Furthermore, we show that federated co-training can collaboratively train interpretable models, such as decision trees and rule ensembles, achieving a model quality comparable to centralized training.
BibTeX
@misc{abourayya2023protecting,
  author = {Amr Abourayya and Jens Kleesiek and Kanishka Rao and Erman Ayday and Bharat Rao and Geoff Webb and Michael Kamp},
  title = {{Protecting Sensitive Data through Federated Co-Training}},
  year = {2023},
  eprint = {2310.05696},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.05696},
}
Figure