Federated learning enhanced by differential privacy has emerged as a popular approach to better safeguard the privacy of client-side data by protecting clients' contributions during the training process. Existing solutions typically assume a uniform privacy budget for all records and provide one-size-fits-all solutions that may not be adequate to meet each record's privacy requirement. In this paper, we explore the uncharted territory of cross-silo FL with record-level personalized differential privacy. We devise a novel framework named rPDP-FL, employing a two-stage hybrid sampling scheme with both client-level sampling and non-uniform record-level sampling to accommodate varying privacy requirements. A critical and non-trivial problem is to select the ideal per-record sampling probability q given the personalized privacy budget {\epsilon}. We introduce a versatile solution named Simulation-CurveFitting, allowing us to uncover a significant insight into the nonlinear correlation between q and {\epsilon} and derive an elegant mathematical model to tackle the problem. Our evaluation demonstrates that our solution can provide significant performance gains over the baselines that do not consider personalized privacy preservation.
Machine unlearning aims to eliminate the influence of a subset of training samples (i.e., unlearning samples) from a trained model. Effectively and efficiently removing the unlearning samples without negatively impacting the overall model performance is still challenging. In this paper, we propose a contrastive unlearning framework, leveraging the concept of representation learning for more effective unlearning. It removes the influence of unlearning samples by contrasting their embeddings against the remaining samples so that they are pushed away from their original classes and pulled toward other classes. By directly optimizing the representation space, it effectively removes the influence of unlearning samples while maintaining the representations learned from the remaining samples. Experiments on a variety of datasets and models on both class unlearning and sample unlearning showed that contrastive unlearning achieves the best unlearning effects and efficiency with the lowest performance loss compared with the state-of-the-art algorithms.
Differential Privacy (DP) was originally developed to protect privacy. However, it has recently been utilized to secure machine learning (ML) models from poisoning attacks, with DP-SGD receiving substantial attention. Nevertheless, a thorough investigation is required to assess the effectiveness of different DP techniques in preventing backdoor attacks in practice. In this paper, we investigate the effectiveness of DP-SGD and, for the first time in literature, examine PATE in the context of backdoor attacks. We also explore the role of different components of DP algorithms in defending against backdoor attacks and will show that PATE is effective against these attacks due to the bagging structure of the teacher models it employs. Our experiments reveal that hyperparameters and the number of backdoors in the training dataset impact the success of DP algorithms. Additionally, we propose Label-DP as a faster and more accurate alternative to DP-SGD and PATE. We conclude that while Label-DP algorithms generally offer weaker privacy protection, accurate hyper-parameter tuning can make them more effective than DP methods in defending against backdoor attacks while maintaining model accuracy.
Differentially Private Federated Learning (DP-FL) has garnered attention as a collaborative machine learning approach that ensures formal privacy. Most DP-FL approaches ensure DP at the record-level within each silo for cross-silo FL. However, a single user's data may extend across multiple silos, and the desired user-level DP guarantee for such a setting remains unknown. In this study, we present ULDP-FL, a novel FL framework designed to guarantee user-level DP in cross-silo FL where a single user's data may belong to multiple silos. Our proposed algorithm directly ensures user-level DP through per-user weighted clipping, departing from group-privacy approaches. We provide a theoretical analysis of the algorithm's privacy and utility. Additionally, we enhance the algorithm's utility and showcase its private implementation using cryptographic building blocks. Empirical experiments on real-world datasets show substantial improvements in our methods in privacy-utility trade-offs under user-level DP compared to baseline methods. To the best of our knowledge, our work is the first FL framework that effectively provides user-level DP in the general cross-silo FL setting.
Federated Learning, as a popular paradigm for collaborative training, is vulnerable against privacy attacks. Different privacy levels regarding users' attitudes need to be satisfied locally, while a strict privacy guarantee for the global model is also required centrally. Personalized Local Differential Privacy (PLDP) is suitable for preserving users' varying local privacy, yet only provides a central privacy guarantee equivalent to the worst-case local privacy level. Thus, achieving strong central privacy as well as personalized local privacy with a utility-promising model is a challenging problem. In this work, a general framework (APES) is built up to strengthen model privacy under personalized local privacy by leveraging the privacy amplification effect of the shuffle model. To tighten the privacy bound, we quantify the heterogeneous contributions to the central privacy user by user. The contributions are characterized by the ability of generating "echos" from the perturbation of each user, which is carefully measured by proposed methods Neighbor Divergence and Clip-Laplace Mechanism. Furthermore, we propose a refined framework (S-APES) with the post-sparsification technique to reduce privacy loss in high-dimension scenarios. To the best of our knowledge, the impact of shuffling on personalized local privacy is considered for the first time. We provide a strong privacy amplification effect, and the bound is tighter than the baseline result based on existing methods for uniform local privacy. Experiments demonstrate that our frameworks ensure comparable or higher accuracy for the global model.
Adversarial examples are crafted by adding indistinguishable perturbations to normal examples in order to fool a well-trained deep learning model to misclassify. In the context of computer vision, this notion of indistinguishability is typically bounded by $L_{\infty}$ or other norms. However, these norms are not appropriate for measuring indistinguishiability for time series data. In this work, we propose adversarial examples in the Wasserstein space for time series data for the first time and utilize Wasserstein distance to bound the perturbation between normal examples and adversarial examples. We introduce Wasserstein projected gradient descent (WPGD), an adversarial attack method for perturbing univariant time series data. We leverage the closed-form solution of Wasserstein distance in the 1D space to calculate the projection step of WPGD efficiently with the gradient descent method. We further propose a two-step projection so that the search of adversarial examples in the Wasserstein space is guided and constrained by Euclidean norms to yield more effective and imperceptible perturbations. We empirically evaluate the proposed attack on several time series datasets in the healthcare domain. Extensive results demonstrate that the Wasserstein attack is powerful and can successfully attack most of the target classifiers with a high attack success rate. To better study the nature of Wasserstein adversarial example, we evaluate a strong defense mechanism named Wasserstein smoothing for potential certified robustness defense. Although the defense can achieve some accuracy gain, it still has limitations in many cases and leaves space for developing a stronger certified robustness method to Wasserstein adversarial examples on univariant time series data.
Deep learning models trained on large-scale data have achieved encouraging performance in many real-world tasks. Meanwhile, publishing those models trained on sensitive datasets, such as medical records, could pose serious privacy concerns. To counter these issues, one of the current state-of-the-art approaches is the Private Aggregation of Teacher Ensembles, or PATE, which achieved promising results in preserving the utility of the model while providing a strong privacy guarantee. PATE combines an ensemble of "teacher models" trained on sensitive data and transfers the knowledge to a "student" model through the noisy aggregation of teachers' votes for labeling unlabeled public data which the student model will be trained on. However, the knowledge or voted labels learned by the student are noisy due to private aggregation. Learning directly from noisy labels can significantly impact the accuracy of the student model. In this paper, we propose the PATE++ mechanism, which combines the current advanced noisy label training mechanisms with the original PATE framework to enhance its accuracy. A novel structure of Generative Adversarial Nets (GANs) is developed in order to integrate them effectively. In addition, we develop a novel noisy label detection mechanism for semi-supervised model training to further improve student model performance when training with noisy labels. We evaluate our method on Fashion-MNIST and SVHN to show the improvements on the original PATE on all measures.
Graph Neural Networks (GNNs) have achieved great success in mining graph-structured data. Despite the superior performance of GNNs in learning graph representations, serious privacy concerns have been raised for the trained models which could expose the sensitive information of graphs. We conduct the first formal study of training GNN models to ensure utility while satisfying the rigorous node-level differential privacy considering the private information of both node features and edges. We adopt the training framework utilizing personalized PageRank to decouple the message-passing process from feature aggregation during training GNN models and propose differentially private PageRank algorithms to protect graph topology information formally. Furthermore, we analyze the privacy degradation caused by the sampling process dependent on the differentially private PageRank results during model training and propose a differentially private GNN (DPGNN) algorithm to further protect node features and achieve rigorous node-level differential privacy. Extensive experiments on real-world graph datasets demonstrate the effectiveness of the proposed algorithms for providing node-level differential privacy while preserving good model utility.
Automatic Speech Recognition models require large amount of speech data for training, and the collection of such data often leads to privacy concerns. Federated learning has been widely used and is considered to be an effective decentralized technique by collaboratively learning a shared prediction model while keeping the data local on different clients devices. However, the limited computation and communication resources on clients devices present practical difficulties for large models. To overcome such challenges, we propose Federated Pruning to train a reduced model under the federated setting, while maintaining similar performance compared to the full model. Moreover, the vast amount of clients data can also be leveraged to improve the pruning results compared to centralized training. We explore different pruning schemes and provide empirical evidence of the effectiveness of our methods.
Tensor factorization has received increasing interest due to its intrinsic ability to capture latent factors in multi-dimensional data with many applications such as recommender systems and Electronic Health Records (EHR) mining. PARAFAC2 and its variants have been proposed to address irregular tensors where one of the tensor modes is not aligned, e.g., different users in recommender systems or patients in EHRs may have different length of records. PARAFAC2 has been successfully applied on EHRs for extracting meaningful medical concepts (phenotypes). Despite recent advancements, current models' predictability and interpretability are not satisfactory, which limits its utility for downstream analysis. In this paper, we propose MULTIPAR: a supervised irregular tensor factorization with multi-task learning. MULTIPAR is flexible to incorporate both static (e.g. in-hospital mortality prediction) and continuous or dynamic (e.g. the need for ventilation) tasks. By supervising the tensor factorization with downstream prediction tasks and leveraging information from multiple related predictive tasks, MULTIPAR can yield not only more meaningful phenotypes but also better predictive performance for downstream tasks. We conduct extensive experiments on two real-world temporal EHR datasets to demonstrate that MULTIPAR is scalable and achieves better tensor fit with more meaningful subgroups and stronger predictive performance compared to existing state-of-the-art methods.