We propose a reparametrization scheme to address the challenges of applying differentially private SGD on large neural networks, which are 1) the huge memory cost of storing individual gradients, 2) the added noise suffering notorious dimensional dependence. Specifically, we reparametrize each weight matrix with two \emph{gradient-carrier} matrices of small dimension and a \emph{residual weight} matrix. We argue that such reparametrization keeps the forward/backward process unchanged while enabling us to compute the projected gradient without computing the gradient itself. To learn with differential privacy, we design \emph{reparametrized gradient perturbation (RGP)} that perturbs the gradients on gradient-carrier matrices and reconstructs an update for the original weight from the noisy gradients. Importantly, we use historical updates to find the gradient-carrier matrices, whose optimality is rigorously justified under linear regression and empirically verified with deep learning tasks. RGP significantly reduces the memory cost and improves the utility. For example, we are the first able to apply differential privacy on the BERT model and achieve an average accuracy of $83.9\%$ on four downstream tasks with $\epsilon=8$, which is within $5\%$ loss compared to the non-private baseline but enjoys much lower privacy leakage risk.
相關內容
Federated learning (FL) emerged as a promising learning paradigm to enable a multitude of participants to construct a joint ML model without exposing their private training data. Existing FL designs have been shown to exhibit vulnerabilities which can be exploited by adversaries both within and outside of the system to compromise data privacy. However, most current works conduct attacks by leveraging gradients on a small batch of data, which is less practical in FL. In this work, we consider a more practical and interesting scenario in which participants share their epoch-averaged gradients (share gradients after at least 1 epoch of local training) rather than per-example or small batch-averaged gradients as in previous works. We perform the first systematic evaluation of attribute reconstruction attack (ARA) launched by the malicious server in the FL system, and empirically demonstrate that the shared epoch-averaged local model gradients can reveal sensitive attributes of local training data of any victim participant. To achieve this goal, we develop a more effective and efficient gradient matching based method called cos-matching to reconstruct the training data attributes. We evaluate our attacks on a variety of real-world datasets, scenarios, assumptions. Our experiments show that our proposed method achieves better attribute attack performance than most existing baselines.
This paper studies the subspace clustering problem in which data points collected from high-dimensional ambient space lie in a union of linear subspaces. Subspace clustering becomes challenging when the dimension of intersection between subspaces is large and most of the self-representation based methods are sensitive to the intersection between the span of clusters. In sharp contrast to the self-representation based methods, a recently proposed clustering method termed Innovation Pursuit, computed a set of optimal directions (directions of innovation) to build the adjacency matrix. This paper focuses on the Innovation Pursuit Algorithm to shed light on its impressive performance when the subspaces are heavily intersected. It is shown that in contrast to most of the existing methods which require the subspaces to be sufficiently incoherent with each other, Innovation Pursuit only requires the innovative components of the subspaces to be sufficiently incoherent with each other. These new sufficient conditions allow the clusters to be strongly close to each other. Motivated by the presented theoretical analysis, a simple yet effective projection based technique is proposed which we show with both numerical and theoretical results that it can boost the performance of Innovation Pursuit.
Although federated learning improves privacy of training data by exchanging local gradients or parameters rather than raw data, the adversary still can leverage local gradients and parameters to obtain local training data by launching reconstruction and membership inference attacks. To defend such privacy attacks, many noises perturbation methods (like differential privacy or CountSketch matrix) have been widely designed. However, the strong defence ability and high learning accuracy of these schemes cannot be ensured at the same time, which will impede the wide application of FL in practice (especially for medical or financial institutions that require both high accuracy and strong privacy guarantee). To overcome this issue, in this paper, we propose \emph{an efficient model perturbation method for federated learning} to defend reconstruction and membership inference attacks launched by curious clients. On the one hand, similar to the differential privacy, our method also selects random numbers as perturbed noises added to the global model parameters, and thus it is very efficient and easy to be integrated in practice. Meanwhile, the random selected noises are positive real numbers and the corresponding value can be arbitrarily large, and thus the strong defence ability can be ensured. On the other hand, unlike differential privacy or other perturbation methods that cannot eliminate the added noises, our method allows the server to recover the true gradients by eliminating the added noises. Therefore, our method does not hinder learning accuracy at all.
Continual Learning (CL) investigates how to train Deep Networks on a stream of tasks without incurring catastrophic forgetting. CL settings proposed in the literature assume that every incoming example is paired with ground-truth annotations. However, this clashes with many real-world applications: gathering labeled data, which is in itself tedious and expensive, becomes indeed infeasible when data flow as a stream and must be consumed in real-time. This work explores Weakly Supervised Continual Learning (WSCL): here, only a small fraction of labeled input examples are shown to the learner. We assess how current CL methods (e.g.: EWC, LwF, iCaRL, ER, GDumb, DER) perform in this novel and challenging scenario, in which overfitting entangles forgetting. Subsequently, we design two novel WSCL methods which exploit metric learning and consistency regularization to leverage unsupervised data while learning. In doing so, we show that not only our proposals exhibit higher flexibility when supervised information is scarce, but also that less than 25% labels can be enough to reach or even outperform SOTA methods trained under full supervision.
Federated learning enables multiple parties to collaboratively train a machine learning model without communicating their local data. A key challenge in federated learning is to handle the heterogeneity of local data distribution across parties. Although many studies have been proposed to address this challenge, we find that they fail to achieve high performance in image datasets with deep learning models. In this paper, we propose MOON: model-contrastive federated learning. MOON is a simple and effective federated learning framework. The key idea of MOON is to utilize the similarity between model representations to correct the local training of individual parties, i.e., conducting contrastive learning in model-level. Our extensive experiments show that MOON significantly outperforms the other state-of-the-art federated learning algorithms on various image classification tasks.
The remarkable practical success of deep learning has revealed some major surprises from a theoretical perspective. In particular, simple gradient methods easily find near-optimal solutions to non-convex optimization problems, and despite giving a near-perfect fit to training data without any explicit effort to control model complexity, these methods exhibit excellent predictive accuracy. We conjecture that specific principles underlie these phenomena: that overparametrization allows gradient methods to find interpolating solutions, that these methods implicitly impose regularization, and that overparametrization leads to benign overfitting. We survey recent theoretical progress that provides examples illustrating these principles in simpler settings. We first review classical uniform convergence results and why they fall short of explaining aspects of the behavior of deep learning methods. We give examples of implicit regularization in simple settings, where gradient methods lead to minimal norm functions that perfectly fit the training data. Then we review prediction methods that exhibit benign overfitting, focusing on regression problems with quadratic loss. For these methods, we can decompose the prediction rule into a simple component that is useful for prediction and a spiky component that is useful for overfitting but, in a favorable setting, does not harm prediction accuracy. We focus specifically on the linear regime for neural networks, where the network can be approximated by a linear model. In this regime, we demonstrate the success of gradient flow, and we consider benign overfitting with two-layer networks, giving an exact asymptotic analysis that precisely demonstrates the impact of overparametrization. We conclude by highlighting the key challenges that arise in extending these insights to realistic deep learning settings.
Federated learning (FL) is an emerging, privacy-preserving machine learning paradigm, drawing tremendous attention in both academia and industry. A unique characteristic of FL is heterogeneity, which resides in the various hardware specifications and dynamic states across the participating devices. Theoretically, heterogeneity can exert a huge influence on the FL training process, e.g., causing a device unavailable for training or unable to upload its model updates. Unfortunately, these impacts have never been systematically studied and quantified in existing FL literature. In this paper, we carry out the first empirical study to characterize the impacts of heterogeneity in FL. We collect large-scale data from 136k smartphones that can faithfully reflect heterogeneity in real-world settings. We also build a heterogeneity-aware FL platform that complies with the standard FL protocol but with heterogeneity in consideration. Based on the data and the platform, we conduct extensive experiments to compare the performance of state-of-the-art FL algorithms under heterogeneity-aware and heterogeneity-unaware settings. Results show that heterogeneity causes non-trivial performance degradation in FL, including up to 9.2% accuracy drop, 2.32x lengthened training time, and undermined fairness. Furthermore, we analyze potential impact factors and find that device failure and participant bias are two potential factors for performance degradation. Our study provides insightful implications for FL practitioners. On the one hand, our findings suggest that FL algorithm designers consider necessary heterogeneity during the evaluation. On the other hand, our findings urge system providers to design specific mechanisms to mitigate the impacts of heterogeneity.
Train machine learning models on sensitive user data has raised increasing privacy concerns in many areas. Federated learning is a popular approach for privacy protection that collects the local gradient information instead of real data. One way to achieve a strict privacy guarantee is to apply local differential privacy into federated learning. However, previous works do not give a practical solution due to three issues. First, the noisy data is close to its original value with high probability, increasing the risk of information exposure. Second, a large variance is introduced to the estimated average, causing poor accuracy. Last, the privacy budget explodes due to the high dimensionality of weights in deep learning models. In this paper, we proposed a novel design of local differential privacy mechanism for federated learning to address the abovementioned issues. It is capable of making the data more distinct from its original value and introducing lower variance. Moreover, the proposed mechanism bypasses the curse of dimensionality by splitting and shuffling model updates. A series of empirical evaluations on three commonly used datasets, MNIST, Fashion-MNIST and CIFAR-10, demonstrate that our solution can not only achieve superior deep learning performance but also provide a strong privacy guarantee at the same time.
We detail a new framework for privacy preserving deep learning and discuss its assets. The framework puts a premium on ownership and secure processing of data and introduces a valuable representation based on chains of commands and tensors. This abstraction allows one to implement complex privacy preserving constructs such as Federated Learning, Secure Multiparty Computation, and Differential Privacy while still exposing a familiar deep learning API to the end-user. We report early results on the Boston Housing and Pima Indian Diabetes datasets. While the privacy features apart from Differential Privacy do not impact the prediction accuracy, the current implementation of the framework introduces a significant overhead in performance, which will be addressed at a later stage of the development. We believe this work is an important milestone introducing the first reliable, general framework for privacy preserving deep learning.
Asynchronous distributed machine learning solutions have proven very effective so far, but always assuming perfectly functioning workers. In practice, some of the workers can however exhibit Byzantine behavior, caused by hardware failures, software bugs, corrupt data, or even malicious attacks. We introduce \emph{Kardam}, the first distributed asynchronous stochastic gradient descent (SGD) algorithm that copes with Byzantine workers. Kardam consists of two complementary components: a filtering and a dampening component. The first is scalar-based and ensures resilience against $\frac{1}{3}$ Byzantine workers. Essentially, this filter leverages the Lipschitzness of cost functions and acts as a self-stabilizer against Byzantine workers that would attempt to corrupt the progress of SGD. The dampening component bounds the convergence rate by adjusting to stale information through a generic gradient weighting scheme. We prove that Kardam guarantees almost sure convergence in the presence of asynchrony and Byzantine behavior, and we derive its convergence rate. We evaluate Kardam on the CIFAR-100 and EMNIST datasets and measure its overhead with respect to non Byzantine-resilient solutions. We empirically show that Kardam does not introduce additional noise to the learning procedure but does induce a slowdown (the cost of Byzantine resilience) that we both theoretically and empirically show to be less than $f/n$, where $f$ is the number of Byzantine failures tolerated and $n$ the total number of workers. Interestingly, we also empirically observe that the dampening component is interesting in its own right for it enables to build an SGD algorithm that outperforms alternative staleness-aware asynchronous competitors in environments with honest workers.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191