Federated Learning (FL) enables collaborative model training across distributed participants without sharing raw data, offering a privacy-preserving paradigm. However, recent studies on gradient inversion attacks have demonstrated the vulnerability of FL to adversaries who can reconstruct sensitive local training data from shared gradients. To mitigate this threat, we propose Gradient Dropout, a novel defense mechanism that disrupts reconstruction attempts while preserving model utility. Specifically, Gradient Dropout perturbs gradients by randomly scaling a subset of components and replacing the remainder with Gaussian noise, thereby creating a transformed gradient space that significantly impedes reconstruction attempts. Moreover, this mechanism is applied across all layers of the model, ensuring that attackers cannot exploit any unperturbed gradients. Theoretical analysis reveals that the perturbed gradients can be kept sufficiently distant from their true values, thereby providing safety guarantees for the proposed algorithm. Furthermore, we demonstrate that this protection mechanism minimally impacts model performance, as gradient dropout and the original training dynamics remain effectively bounded under certain convexity conditions. These findings are substantiated through experimental evaluations, where we show that various attack methods yield low-quality reconstructed images while model performance is largely preserved, with less than 2% accuracy reduction relative to the baseline. As such, Gradient Dropout is presented as an effective solution for safeguarding privacy in FL, providing a balanced trade-off between privacy protection, computational efficiency, and model accuracy.

In this article, we propose a method, generative image reconstruction from gradients (GIRG), for recovering training images from gradients in a federated learning (FL) setting, where privacy is preserved by sharing model weights and gradients rather than raw training data. Previous studies have shown the potential for revealing clients' private information or even pixel-level recovery of training images from shared gradients. However, existing methods are limited to low-resolution images and small batch sizes (BSs) or require prior knowledge about the client data. GIRG utilizes a conditional generative model to reconstruct training images and their corresponding labels from the shared gradients. Unlike previous generative model-based methods, GIRG does not require prior knowledge of the training data. Furthermore, GIRG optimizes the weights of the conditional generative model to generate highly accurate "dummy" images instead of optimizing the input vectors of the generative model. Comprehensive empirical results show that GIRG is able to recover high-resolution images with large BSs and can even recover images from the aggregation of gradients from multiple participants. These results reveal the vulnerability of current FL practices and call for immediate efforts to prevent inversion attacks in gradient-sharing-based collaborative training.

In contrast to the prevailing interpretation of Low-Rank Adaptation (LoRA) as a means of simulating weight changes in model adaptation, this paper introduces an alternative perspective by framing it as a control process. Specifically, we conceptualize lightweight matrices in LoRA as control modules tasked with perturbing the original, complex, yet frozen blocks on downstream tasks. Building upon this new understanding, we conduct a thorough analysis on the controllability of these modules, where we identify and establish sufficient conditions that facilitate their effective integration into downstream controls. Moreover, the control modules are redesigned by incorporating nonlinearities through a parameter-free attention mechanism. This modification allows for the intermingling of tokens within the controllers, enhancing the adaptability and performance of the system. Empirical findings substantiate that, without introducing any additional parameters, this approach surpasses the existing LoRA algorithms across all assessed datasets and rank configurations.

Multi-Party Computation (MPC) provides an effective cryptographic solution for distributed computing systems so that local models with sensitive information are encrypted before sending to the centralized servers for aggregation. Though direct local knowledge leakages are eliminated in MPC-based algorithms, we observe the server can still obtain the local information indirectly in many scenarios, or even reveal the groundtruth images through methods like Deep Leakage from Gradients (DLG). To eliminate such possibilities and provide stronger protections, we propose an augmented MPC approach by encrypting local models with two rounds of decomposition before transmitting to the server. The proposed solution allows us to remove the constraint that servers must be honest in the general federated learning settings since the true global model is hidden from the servers. Specifically, the augmented MPC algorithm encodes local models into multiple secret shares in the first round, then each share is furthermore split into a public share and a private share. Consequences of such a two-round decomposition are that the augmented algorithm fully inherits the advantages of standard MPC by providing lossless encryption and decryption while simultaneously rendering the global model invisible to the central server. Both theoretical analysis and experimental verification demonstrate that such an augmented solution can provide stronger protections for the security and privacy of the training data, with minimal extra communication and computation costs incurred.

Federated learning has gained recognitions as a secure approach for safeguarding local private data in collaborative learning. But the advent of gradient inversion research has posed significant challenges to this premise by enabling a third-party to recover groundtruth images via gradients. While prior research has predominantly focused on low-resolution images and small batch sizes, this study highlights the feasibility of reconstructing complex images with high resolutions and large batch sizes. The success of the proposed method is contingent on constructing an over-parameterized convolutional network, so that images are generated before fitting to the gradient matching requirement. Practical experiments demonstrate that the proposed algorithm achieves high-fidelity image recovery, surpassing state-of-the-art competitors that commonly fail in more intricate scenarios. Consequently, our study shows that local participants in a federated learning system are vulnerable to potential data leakage issues. Source code is available at https://github.com/czhang024/CI-Net.

Multi-party computation (MPC) allows distributed machine learning to be performed in a privacy-preserving manner so that end-hosts are unaware of the true models on the clients. However, the standard MPC algorithm also triggers additional communication and computation costs, due to those expensive cryptography operations and protocols. In this paper, instead of applying heavy MPC over the entire local models for secure model aggregation, we propose to encrypt critical part of model (gradients) parameters to reduce communication cost, while maintaining MPC's advantages on privacy-preserving without sacrificing accuracy of the learnt joint model. Theoretical analysis and experimental results are provided to verify that our proposed method could prevent deep leakage from gradients attacks from reconstructing original data of individual participants. Experiments using deep learning models over the MNIST and CIFAR-10 datasets empirically demonstrate that our proposed partially encrypted MPC method can reduce the communication and computation cost significantly when compared with conventional MPC, and it achieves as high accuracy as traditional distributed learning which aggregates local models using plain text.

In this paper, we propose a decentralized distributed algorithm with stochastic communication among nodes, building on a sampling method called "edge sampling''. Such a sampling algorithm allows us to avoid the heavy peer-to-peer communication cost when combining neighboring weights on dense networks while still maintains a comparable convergence rate. In particular, we quantitatively analyze its theoretical convergence properties, as well as the optimal sampling rate over the underlying network. When compared with previous methods, our solution is shown to be unbiased, communication-efficient and suffers from lower sampling variances. These theoretical findings are validated by both numerical experiments on the mixing rates of Markov Chains and distributed machine learning problems.

The goal of online active learning is to learn predictive models from a sequence of unlabeled data given limited label query budget. Unlike conventional online learning tasks, online active learning is considerably more challenging because of two reasons. First, it is difficult to design an effective query strategy to decide when is appropriate to query the label of an incoming instance given limited query budget. Second, it is also challenging to decide how to update the predictive models effectively whenever the true label of an instance is queried. Most existing approaches for online active learning are often based on a family of first-order online learning algorithms, which are simple and efficient but fall short in the slow convergence and sub-optimal solution in exploiting the labeled training data. To solve these issues, this paper presents a novel framework of Second-order Online Active Learning (SOAL) by fully exploiting both the first-order and second-order information. The proposed algorithms are able to achieve effective online learning efficacy, maximize the predictive accuracy, and minimize the labeling cost. To make SOAL more practical for real-world applications, especially for class-imbalanced online classification tasks (e.g., malicious web detection), we extend the SOAL framework by proposing the Costsensitive Second-order Online Active Learning algorithm named “SOALCS”, which is devised by maximizing the sum of weighted sensitivity and specificity or minimizing the cost of weighted mistakes of different classes. We conducted both theoretical analysis and empirical studies, including an extensive set of experiments on a variety of large-scale real-world datasets, in which the promising empirical results validate the efficacy and scalability of the proposed algorithms towards large-scale online learning tasks.

This paper investigates the problem of online active learning for training classification models from sequentially arriving data. This is more challenging than conventional online learning tasks since the learner not only needs to figure out how to effectively update the classifier but also needs to decide when is the best time to query the label of an incoming instance given limited label budget. The existing online active learning approaches are often based on first-order online learning methods which generally fall short in slow convergence rate and sub-optimal exploitation of available information when querying the labeled data. To overcome the limitations, in this paper, we present a new framework of Second-order Online Active Learning (SOAL), which fully exploits both first-order and second-order information to achieve high learning accuracy with low labeling cost. We conduct both theoretical analysis and empirical studies for evaluating the proposed SOAL algorithm extensively. The encouraging results show clear advantages of the proposed algorithm over a family of state-of-the-art online active learning algorithms.

A series of online multi-task learning (OMTL) algorithms have been proposed to avoid the expensive training cost and poor adaptability of traditional batch multi-task learning (MTL) algorithms in recent years. However, these OMTL algorithms usually assume that all tasks are closely related, which may not hold in practical scenarios. More importantly, their theoretical reliability is weakened due to the lack of proof on the cumulative regrets. To overcome these limitations, we present a robust online multi-task classification framework (ROM) and its two optimization algorithms (ROM-PGD, ROM-RDA). The proposed algorithms can not only automatically capture the common features among all tasks and individual features for each task, but also identify the potential existence of outlier task. Theoretically, we prove that the regret bounds of these two algorithms are sub-linear compared with the best separating algorithm in hindsight. Empirical studies on both synthetic and real-world datasets also demonstrate the effectiveness of our proposed algorithms when compared with the state-of-the-art OMTL algorithms.