Federated unlearning (FU) allows a participating client in a federated learning (FL) system to remove its contribution from the trained global model, thereby enforcing the client’s ``right to be forgotten'' (RTBF). However, from the perspective of a client that does not request unlearning, the activation of the FU process may disrupt ongoing FL training and introduce additional computational and time overhead. In such cases, a client opposed to unlearning may be incentivized to retaliate against the unlearning client(s). In this work, we take the first step toward demonstrating the feasibility of such retaliatory behavior by exploiting the information leakage introduced during the FU process. Specifically, we propose a novel unlearning-induced membership inference attack (MIA) model, followed by a coarse-to-fine data generation method that enables an adversarial client to locally reconstruct the unlearned data. Building on this reconstruction, we introduce two targeted retaliatory attacks: (1) Anti-Unlearning Attack (AUA), which hinders the global model from successfully forgetting the data intended for removal, and (2) Discrimination-Unlearning Attack (DUA), which specifically degrades the global model’s performance on the unlearned data. Extensive experiments across a variety of FU methods and settings validate the effectiveness of the proposed retaliatory attack framework.

How to defend against attacks in Federated Partial Label Learning (FedPLL) is a brand new and challenging question in machine learning security due to stealthy and efficient attack behaviors of adversaries. In this paper, we systematically study this problem by developing an Adaptive Partial Label Attack (APLA) which subtly manipulates the candidate label set of the data sample. To defend against APLA, we develop the RobustFedPLL framework incorporating three modules: (1) in preliminary clustering, we implement a Gaussian Mixture Model (GMM) and a moving average mechanism to identify clients' confidence; (2) in representation contrasting, we develop a contrast-based algorithm to obtain clients' model feature representations; (3) in final clustering, we utilize mainstream clustering algorithms to finally distinguish adversaries. Experiments of RobustFedPLL and SOTA defense algorithms based on two datasets are conducted, demonstrating the superiority of RobustFedPLL under various experimental settings.

In this paper, we propose pFedMo, a personalized federated learning algorithm with contrastive momentum. In pFedMo, we design a score function to personalize worker models by distilling knowledge from the aggregator's representation model so as to address the non-i.i.d. issue. To accelerate the convergence, we leverage the momentum acceleration on both the worker side and the aggregator side. However, the typical momentum without personalization does not suit well for the worker models with personalization, influencing convergence performance. To address this, we develop a personalized/contrastive momentum method for efficient momentum acceleration. We provide mathematical proof for the convergence of pFedMo on non-i.i.d. data. Extensive experiments based on real-world datasets and IoT system are conducted, verifying that pFedMo outperforms existing mainstream benchmarks, and achieves up to 35.90% accuracy increase and 3.64x training time speedup under a wide range of settings.

The problem of missing data in the clinical environment is a common one. However, the presence of these missing values or incorrect processing can affect downstream tasks. Imputation is the most effective way to solve the missing problem. Combining deep generative models and multiple imputation idea provides new ideas for the imputation of data, but some methods do not apply to mixed-type tabular data such as clinical data. The missing types of tabular data in real-world scenarios are varied and unpredictable, so it is necessary to construct a generalized imputation method. In this paper, we propose a multiple imputation architecture for missing neighborhood perturbation denoising autoencoder (MI_NPDAE) for imputing mixed types of incomplete medical tabular data for different missing mechanisms. In MI_NPDAE, the neighborhood information of the data is searched to find the best set of donors to be used as the input layer of the network, and the learning of the neighborhood features of the missing locations is ensured by reconstructing new inputs, while the addition of additive noise exposes the model to different missing mechanisms. The experiments were respectively conducted on a public dataset Breast and one lung cancer nutrition dataset from the Chinese Anti-Cancer Society. The experiments on the datasets show that the proposed model shows better results under different missing mechanisms and can maintain a low error on the comparison experiments on data with different missing ratios. In addition, the imputed results also show significant advantages in downstream prediction tasks.

Accurate clinical cancer prediction is important for clinical diagnosis. Most of the cancer patient data have high privacy and using clinical patient data directly may create concerns such as privacy leakage, which can be solved by replacing real data with synthetic data for clinical trial modelling. Synthetic data can also fill in the missing, unbalanced clinical datasets. Data augmentation can achieve data synthesis. However, traditional data augmentation models are mostly based on convolutional neural network architectures, which can only capture local dependencies and do not well describe the feature-related information of patients over long survival times. In this paper, we propose a self-attention medical tabular data augmentation model with privacy protection, ACCT-GAN, based on the generative adversarial network model, which uses the AC block generated by combining the ACmix model to reconstruct the discriminator model in the CTAB-GAN+ network, which can effectively solve the limitations of the original convolutional neural network, model the global relationship, and make the generated data retain the feature correlation between the original data better correlations between the original data. Meanwhile, taking into account the privacy of medical data, the DP-SGD algorithm of CTAB-GAN+ is optimised to reduce the privacy risk of synthetic data. The results show that ACCT-GAN synthesizes privacy-preserving data with at least 26.21% higher utility across the dateset of nasopharyngeal cancer patients provided by Xiangya Medical College and learning tasks under different privacy budgets, demonstrating the usability, high quality, and better privacy preservation of the generated data, validating the effectiveness of this paper’s method, and confirming the potential of this paper’s method for cancer datasets.

In this paper, we propose and analyze HierAdMo, a three-tier adaptive momentum accelerated client-edge-cloud Federated Learning (FL) algorithm. HierAdMo combines the momentum acceleration on both worker and edge levels. However, simply combining these two levels of momenta may lead to disagreement between them, negatively influencing convergence performance. To this end, we embed an online adaptive method that scales down the momentum when disagreement occurs. We provide mathematical proof for the convergence of HierAdMo for non-i.i.d. data and the tighter convergence upper bound compared with a version of HierAdMo without adaptation (HierAdMo-R). Finally, extensive experiments based on real-world datasets are conducted, verifying that HierAdMo outperforms existing mainstream benchmarks and achieves the optimal or near-optimal convergence performance compared with HierAdMo-R under a wide range of settings.

In this article, we propose Hierarchical Federated Learning with Momentum Acceleration (HierMo), a three-tier worker-edge-cloud federated learning algorithm that applies momentum for training acceleration. Momentum is calculated and aggregated in the three tiers. We provide convergence analysis for HierMo, showing a convergence rate of$\mathcal {O}(\frac{1}{T})$. In the analysis, we develop a new approach to characterize model aggregation, momentum aggregation, and their interactions. Based on this result, we prove that HierMo achieves a tighter convergence upper bound compared with HierFAVG without momentum. We also propose HierOPT, which optimizes the aggregation periods (worker-edge and edge-cloud aggregation periods) to minimize the loss given a limited training time. By conducting the experiment, we verify that HierMo outperforms existing mainstream benchmarks under a wide range of settings. In addition, HierOPT can achieve a near-optimal performance when we test HierMo under different aggregation periods.