End-to-End Differential Privacy in Training Deep Neural Network Classifiers
Authors: Huaiyuan Rao, Calvin Hawkins, Alexander Benvenuti, Matthew Hale
Abstract
Differentially private machine learning enables model training on sensitive data while ensuring that individual data is unlikely to be recoverable from the parameters of the resulting model. However, existing work often privatizes both training inputs and their labels, and these protections may be conservative when labels are public or can be safely made public. Therefore, in this work we propose a novel private training framework that instead privatizes training inputs while keeping labels public. We consider neural networks with softmax output layers, and thus the mapping from training inputs to the output of the softmax layer is a mapping onto the unit simplex. We randomize softmax outputs during training by applying the Dirichlet mechanism to enforce differential privacy for the training inputs, hence the ``end-to-end'' label. Because training data is reused across multiple training epochs, we use the notion of \Renyi differential privacy to formulate tight bounds on the strength of privacy provided by the Dirichlet mechanism across repeated uses. We show empirically that we attain new state-of-the-art accuracy when training from scratch on CIFAR10, MNIST, MedMNIST, FashionMNIST, and SVHN across all privacy budgets evaluated. Notably, when implementing (ε,δ)-differential privacy with δ=10−5, we improve the prior state-of-the-art accuracy from 78.37% to 88.17% at ε=4 on CIFAR10, and our approach has 82.96% accuracy even for ε=1, which significantly outperforms prior work.
Differentially private (DP) training of neural networks is often hindered by the large amount of noise required by gradient-based methods such as DP-SGD, which repeatedly inject high-dimensional noise in parameter space throughout training. In this paper, we propose a new framework for DP learning that avoids iterative optimization in parameter space. Instead of updating the target model using privatized gradients, we employ a hypernetwork trained on public datasets to map a private dataset to the parameters of the target model. Specifically, each example is embedded into a low-dimensional representation, the embeddings are aggregated and perturbed to obtain a DP dataset embedding, and the hypernetwork generates the target model parameters from this noisy embedding. Because privacy noise is injected only once into a low-dimensional dataset representation, our approach can significantly reduce the adverse effect of noise. We theoretically show in a synthetic setting that, under a fixed privacy budget, models produced by our approach achieve higher utility than those trained with DP-SGD. Moreover, we apply our approach to LoRA fine-tuning of diffusion models and show that it achieves lower FID than LoRA models trained with DP-SGD and other public-data-guided methods.
Test-time adaptation (TTA) can reduce error on new and different data by updating the model on these inputs during inference. However, these updates raise the issue of privacy w.r.t. the testing data, because the model parameters now depend on all past inputs. To control this privacy risk, we cast multiple popular TTA methods (Tent, EATA, SAR, DeYO, and COME) into differential privacy (DP) forms that apply per-sample gradient clipping and Gaussian noise for all updates. On ImageNet-C, our DP-TTA methods provide adequate privacy at small cost to accuracy, and in the low-privacy regime the clipping mechanism of DP can even improve the accuracy and stability of adaptation in the continual setting. These improvements to privacy and accuracy come at only modest computational overhead. These first results on private TTA raise awareness of the issue, inform the development of more private test-time updates, and identify per-sample clipping as an effective technique for improving the accuracy and stability of adaptation.
This paper explores the use of linear aggregation to protect the privacy of sensitive training labels through the concept of \emph{label differential privacy} (label-DP) while maintaining regression task utility. Our key finding is that weighted linear aggregation of training instances with i.i.d. N(0,1) weights can achieve (ε,δ)-label-DP with m=O(n/(log(1/δ))). Unlike prior methods, our approach relies on the minimum linear regression loss rather than the minimum singular value of the data matrix, resulting in better practical bounds on real datasets. We also examine real-world mechanisms involving disjoint sets or \textit{bags} of instances. We demonstrate that aggregating labels from sub-sampled disjoint k-sized bags using i.i.d. N(0,1) weights achieves (ε,δ)-label-DP with k≥Ω(((1/ε)log(1/δ))2). In both scenarios, the optimal linear mse-regressor on the aggregated data approximates the original dataset's optimum with high probability, without needing additive label noise. Furthermore, we show that adding N(0,1) noise to any constant fraction of labels allows for similar label-DP guarantees when aggregating labels over random disjoint bags, while preserving the utility of Lipschitz-bounded neural mse-regression tasks.