Parametric Feature Transfer: One-shot Federated Learning with Foundation Models

In one-shot federated learning (FL), clients collaboratively train a global model in a single round of communication. Existing approaches for one-shot FL enhance communication efficiency at the expense of diminished accuracy. This paper introduces FedPFT (Federated Learning with Parametric Feature Transfer), a methodology that harnesses the transferability of foundation models to enhance both accuracy and communication efficiency in one-shot FL. The approach involves transferring per-client parametric models (specifically, Gaussian mixtures) of features extracted from foundation models. Subsequently, each parametric model is employed to generate synthetic features for training a classifier head. Experimental results on eight datasets demonstrate that FedPFT enhances the communication-accuracy frontier in both centralized and decentralized FL scenarios, as well as across diverse data-heterogeneity settings such as covariate shift and task shift, with improvements of up to 20.6%. Additionally, FedPFT adheres to the data minimization principle of FL, as clients do not send real features. We demonstrate that sending real features is vulnerable to potent reconstruction attacks. Moreover, we show that FedPFT is amenable to formal privacy guarantees via differential privacy, demonstrating favourable privacy-accuracy tradeoffs.

Normalization Is All You Need: Understanding Layer-Normalized Federated Learning under Extreme Label Shift

Layer normalization (LN) is a widely adopted deep learning technique especially in the era of foundation models. Recently, LN has been shown to be surprisingly effective in federated learning (FL) with non-i.i.d. data. However, exactly why and how it works remains mysterious. In this work, we reveal the profound connection between layer normalization and the label shift problem in federated learning. To understand layer normalization better in FL, we identify the key contributing mechanism of normalization methods in FL, called feature normalization (FN), which applies normalization to the latent feature representation before the classifier head. Although LN and FN do not improve expressive power, they control feature collapse and local overfitting to heavily skewed datasets, and thus accelerates global training. Empirically, we show that normalization leads to drastic improvements on standard benchmarks under extreme label shift. Moreover, we conduct extensive ablation studies to understand the critical factors of layer normalization in FL. Our results verify that FN is an essential ingredient inside LN to significantly improve the convergence of FL while remaining robust to learning rate choices, especially under extreme label shift where each client has access to few classes.

Private Distribution Learning with Public Data: The View from Sample Compression

We study the problem of private distribution learning with access to public data. In this setup, which we refer to as public-private learning, the learner is given public and private samples drawn from an unknown distribution $p$ belonging to a class $\mathcal Q$, with the goal of outputting an estimate of $p$ while adhering to privacy constraints (here, pure differential privacy) only with respect to the private samples. We show that the public-private learnability of a class $\mathcal Q$ is connected to the existence of a sample compression scheme for $\mathcal Q$, as well as to an intermediate notion we refer to as list learning. Leveraging this connection: (1) approximately recovers previous results on Gaussians over $\mathbb R^d$; and (2) leads to new ones, including sample complexity upper bounds for arbitrary $k$-mixtures of Gaussians over $\mathbb R^d$, results for agnostic and distribution-shift resistant learners, as well as closure properties for public-private learnability under taking mixtures and products of distributions. Finally, via the connection to list learning, we show that for Gaussians in $\mathbb R^d$, at least $d$ public samples are necessary for private learnability, which is close to the known upper bound of $d+1$ public samples.

Private GANs, Revisited

We show that the canonical approach for training differentially private GANs -- updating the discriminator with differentially private stochastic gradient descent (DPSGD) -- can yield significantly improved results after modifications to training. Existing instantiations of this approach neglect to consider how adding noise only to discriminator updates disrupts the careful balance between the generator and discriminator necessary for successful GAN training. We show that a simple fix -- taking more discriminator steps between generator steps -- restores parity and improves results. Additionally, with the goal of restoring parity between the generator and discriminator, we experiment with other modifications to improve discriminator training and see further improvements in generation quality. Our results demonstrate that on standard benchmarks, DPSGD outperforms all alternative GAN privatization schemes.

Private Estimation with Public Data

We initiate the study of differentially private (DP) estimation with access to a small amount of public data. For private estimation of d-dimensional Gaussians, we assume that the public data comes from a Gaussian that may have vanishing similarity in total variation distance with the underlying Gaussian of the private data. We show that under the constraints of pure or concentrated DP, d+1 public data samples are sufficient to remove any dependence on the range parameters of the private data distribution from the private sample complexity, which is known to be otherwise necessary without public data. For separated Gaussian mixtures, we assume that the underlying public and private distributions are the same, and we consider two settings: (1) when given a dimension-independent amount of public data, the private sample complexity can be improved polynomially in terms of the number of mixture components, and any dependence on the range parameters of the distribution can be removed in the approximate DP case; (2) when given an amount of public data linear in the dimension, the private sample complexity can be made independent of range parameters even under concentrated DP, and additional improvements can be made to the overall sample complexity.

Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence

Although machine learning models trained on massive data have led to break-throughs in several areas, their deployment in privacy-sensitive domains remains limited due to restricted access to data. Generative models trained with privacy constraints on private data can sidestep this challenge, providing indirect access to private data instead. We propose DP-Sinkhorn, a novel optimal transport-based generative method for learning data distributions from private data with differential privacy. DP-Sinkhorn minimizes the Sinkhorn divergence, a computationally efficient approximation to the exact optimal transport distance, between the model and data in a differentially private manner and uses a novel technique for control-ling the bias-variance trade-off of gradient estimates. Unlike existing approaches for training differentially private generative models, which are mostly based on generative adversarial networks, we do not rely on adversarial objectives, which are notoriously difficult to optimize, especially in the presence of noise imposed by privacy constraints. Hence, DP-Sinkhorn is easy to train and deploy. Experimentally, we improve upon the state-of-the-art on multiple image modeling benchmarks and show differentially private synthesis of informative RGB images. Project page:https://nv-tlabs.github.io/DP-Sinkhorn.

Fully Quantizing a Simplified Transformer for End-to-end Speech Recognition

While significant improvements have been made in recent years in terms of end-to-end automatic speech recognition (ASR) performance, such improvements were obtained through the use of very large neural networks, unfit for embedded use on edge devices. That being said, in this paper, we work on simplifying and compressing Transformer-based encoder-decoder architectures for the end-to-end ASR task. We empirically introduce a more compact Speech-Transformer by investigating the impact of discarding particular modules on the performance of the model. Moreover, we evaluate reducing the numerical precision of our network's weights and activations while maintaining the performance of the full-precision model. Our experiments show that we can reduce the number of parameters of the full-precision model and then further compress the model 4x by fully quantizing to 8-bit fixed point precision.