Learning Model-Based Privacy Protection under Budget Constraints

Junyuan Hong, Haotao Wang, Zhangyang Wang, Jiayu Zhou

Abstract

Differential Privacy (DP) often introduces a significant "utility tax" due to the noise required to protect sensitive data. Traditional methods rely on hand-tuned noise schedules that are often sub-optimal and difficult to generalize. We present Learning-to-Protect (L2P), a meta-learning framework that automates the design of privacy protection strategies. L2P employs a model-based "protector" consisting of two Recurrent Neural Networks (RNNs) to adaptively allocate privacy budgets and denoise gradients. By meta-training on non-private auxiliary data, L2P learns to optimize the utility-privacy trade-off, significantly outperforming state-of-the-art baselines in high-privacy regimes.

Introduction

Training machine learning models with Differential Privacy involves a fundamental trade-off between model utility and data privacy. The privacy budget, $\epsilon$, acts as a finite resource that is depleted as noise is added to gradients during training. Existing approaches typically use fixed heuristics or simple adaptive rules for noise injection. However, these methods fail to account for the complex dynamics of the optimization landscape and often require manual tuning that might itself leak private information. This work introduces L2P, which treats the discovery of optimal privacy strategies as a learning problem, leveraging auxiliary data to pre-train a protector model.

Methodology

The L2P framework introduces a "Protector" module that replaces standard noise injection mechanisms.

The Model-Based Protector

The protector is composed of two specialized RNNs:

  1. The Scheduler ($\sigma$): This component monitors the training state and dynamically determines the noise level for each gradient update. It learns to strategically preserve the privacy budget for the most impactful training phases.
  2. The Projector ($\pi$): After noise is added, the projector RNN processes the noisy gradient to predict an optimal parameter update. It effectively learns to "denoise" the signal, extracting meaningful patterns even in high-noise environments.
The L2P framework architecture, featuring the Protector module with Scheduler and Projector RNNs.

Privacy Accounting and Meta-Optimization

To facilitate efficient training, the authors propose Constrained Truncated Concentrated DP (ctCDP), a metric that allows for linear accumulation of privacy costs. The protector is meta-trained using a Lagrangian method on public auxiliary datasets. The objective is to minimize the final model loss while strictly adhering to a predefined privacy budget constraint.

Experimental Evaluation

L2P was evaluated on diverse tasks, including logistic regression on IPUMS census data and Multi-Layer Perceptrons (MLP) on MNIST35. A key focus was generalization: training the protector on one dataset (e.g., Brazilian census) and testing it on another (e.g., US census).

Results

L2P consistently outperformed baselines such as Adaptive Gradient Descent (AGD) and Moments Accountant (SGD-MA), particularly in low-budget (high-privacy) scenarios.

Comparison of L2P against various baselines across different datasets and privacy budgets (epsilon). L2P shows superior accuracy and lower loss in high-privacy regimes.

Key findings include:

  • Enhanced Utility: L2P achieves higher accuracy by learning to filter noise through its projector RNN.
  • Adaptive Strategies: The learned schedules often prioritize noise reduction in later stages of training to ensure convergence.
  • Robust Generalization: Protectors trained on auxiliary tasks successfully transfer their strategies to new, private datasets, demonstrating that privacy-preserving patterns are learnable across domains.

Conclusion

Learning-to-Protect (L2P) represents a shift from manual to automated privacy engineering. By framing privacy protection as a meta-learning task, L2P discovers adaptive strategies that significantly improve the utility of DP-trained models. While dependent on the availability of relevant auxiliary data, L2P provides a powerful tool for deploying high-utility models under strict privacy constraints. Future work will explore private meta-optimization to reduce reliance on public datasets.

Summary updated

Abstract

Protecting privacy in gradient-based learning has become increasingly critical as more sensitive information is being used. Many existing solutions seek to protect the sensitive gradients by constraining the overall privacy cost within a constant budget, where the protection is hand-designed and empirically calibrated to boost the utility of the resulting model. However, it remains challenging to choose the proper protection adapted for specific constraints so that the utility is maximized. To this end, we propose a novel Learning-to-Protect algorithm that automatically learns a model-based protector from a set of non-private learning tasks. The learned protector can be applied to private learning tasks to improve utility within the specific privacy budget constraint. Our empirical studies on both synthetic and real datasets demonstrate that the proposed algorithm can achieve a superior utility with a given privacy constraint and generalize well to new private datasets distributed differently as compared to the hand-designed competitors.
BibTeX
@inproceedings{hong2021learning,
  author = {Junyuan Hong and Haotao Wang and Zhangyang Wang and Jiayu Zhou},
  title = {{Learning Model-Based Privacy Protection under Budget Constraints}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i9.16941},
}
Figure