The proliferation of Edge Intelligence (EI) and diverse user demands has led to the generation of vast amounts of heterogeneous multimodal data at the network edge. Multimodal Federated Learning (MFL) offers a promising solution for intelligent and personalized services by enabling collaborative training across distributed clients while preserving data privacy. However, existing MFL frameworks remain unsuitable for edge deployment, as they primarily assume homogeneous environments and fail to address heterogeneous client resources. To bridge this gap, we propose$M$ultimodal$F$ederated$E$dge$L$earning (MFEL), a novel paradigm that extends the conventional MFL framework to enable adaptive submodel deployment based on client capabilities. Building on MFEL, we propose MFEL-HAM, a heterogeneous-aware MFL approach that incorporates three core mechanisms: (1) Prototype Networks to align cross-client modality-specific representations, mitigating divergences caused by non-IID data and heterogeneous sensing environments; (2) Rebalanced Modality Gradient Modulation (R-MGM), which adaptively amplifies gradients of underrepresented modalities and suppresses those of dominant ones, alleviating intra-client modality imbalance; and (3) Momentum Knowledge Distillation (MKD), enabling efficient knowledge transfer without sharing raw data, effectively mitigating the impact of resource heterogeneity on collaborative training. Extensive experiments on heterogeneous multimodal datasets show that MFEL-HAM consistently outperforms baselines in accuracy, convergence speed, and training stability, while demonstrating strong generalization across diverse architectures and resource profiles.

Advances in hardware and AI have enabled edgebased IoT devices to leverage substantial computational and data resources, facilitating large-scale deployment of AI models, particularly through federated learning (FL). However, the high heterogeneity of devices and resource contention at the edge make collaborative optimization of resource scheduling for multiple FL tasks challenging. To tackle this, we propose a novel Multi-Task Asynchronous Federated Learning (MTAFL) architecture, which enhances resource utilization efficiency by enabling orthogonal multiplexing of computation and communication resources through adjusting local epochs on edge devices. Then, we formulate an optimization problem in the MTAFL framework to manage resources and local epochs, aiming to minimize energy consumption while achieving FL performance. However, intricate couplings between resource allocation and local control complicate the long-term FL process. To address this, we employ a two-step relaxation approach and develop an efficient optimization strategy based on the block coordinate descent algorithm. To enhance optimization granularity, we extend the MTAFL framework by incorporating device-level data characteristics. We propose a Gaussian Process-based client selection mechanism that dynamically characterizes and predicts training loss trajectories across clients. After selecting clients for each task, we optimize resource allocation and local control strategies in the system. Extensive numerical evaluations corroborate the superior performance of the proposed approaches over existing schemes.

Local differential privacy (LDP) mechanisms are widely used to collect data generated by IoT sensor devices to protect sensitive information. However, it easily leads to low data utility and high data computing cost due to complex structure and high dimensionality of IoT data. To alleviate this problem, we propose a High-dimensional Data Publishing method using Random responses based on Markov network (HDPRM). This method efficiently conducts the collection and analysis of high-dimensional data under the IoT architecture and satisfies LDP. In particular, it uses the expectation maximization (EM) algorithm to reconstruct the joint distribution of high-dimensional data attributes. Specifically, to improve the effectiveness of data release, we calculate the correlation between attributes and construct corresponding Markov network on the server. Additionally, we cluster high-dimensional attributes through the junction tree algorithm, filter out the joint probability that meets the requirements, and then use this probability to generate synthetic data for publication from the sampled data. Extensive experiments are conducted to comprehensively evaluate the performance of the HDPRM on three real-world datasets. The results show that the method achieves higher data utility under LDP guarantee compared to state-of-the-art methods.

Data privacy protection legislation around the world has increasingly enforced the “right to be forgotten” regulation, generating a surge in research interest in machine unlearning (MU), which aims to remove the impact of training data from machine learning models upon receiving revocation requests from data owners. There exist two major challenges for the performance of MU: the execution efficiency and the inference interference. The former requires minimizing the computational overhead for each execution of the MU mechanism, while the latter calls for reducing the execution frequency to minimize interference with normal inference services. Nowadays most MU studies focus on the sample-level unlearning setting, leaving the other paramount feature-level setting under-explored. Adapting these existing techniques to the latter turns out to be non-trivial. The only known feature-level work achieves anapproximateunlearning guarantee, but suffers from degraded model accuracy and still leaves the inference interference challenge unsolved. We are therefore motivated to propose FELEMN, the first FEature-Level Exact Machine uNlearning method that overcomes both of the above-mentioned hurdles. For the MU execution efficiency challenge, we explore the impact of different feature partitioning strategies on the preservation of semantic relationships for maintaining model accuracy and MU efficiency. For the inference interference challenge, we propose two batching mechanisms to combine as many individual unlearning requests to be processed together as possible, while avoiding potential privacy issues coming with falsely postponing unlearning requests, which is grounded on theoretical analysis. Experiments on five real datasets show that our FELEMN outperforms up-to-date competitors with up to$3\times$speedup for each MU execution, and 50% runtime reduction by mitigating inference interference.

We study online data analytics with differential privacy (DP) in decentralized settings. Specifically, online data analytics with local DP protection is widely adopted in real-world applications. Despite numerous endeavors in this field, significant gaps in utility and functionality remain when compared to its offline counterpart. We present an optimal, streamable mechanism:ExSub, for local DP sparse vector estimation. The mechanism enables a range of online analytics on streaming binary vectors, including multi-dimensional binary, categorical, or set-valued data. By leveraging the negative correlation of occurrence events in the sparse vector, we attain an optimal error rate under local privacy constraints, only requiring streamable computations. To surpass the error barrier of local privacy, we also studyExSubrandomizer in the newly emerging (single-message) shuffle model of DP, and provide nearly-tight privacy amplification bounds therein. Additionally, we leverage the online shuffle model that independently permutes users' messages at each timestamp, to design a simplified randomization strategy that can approximately reach Gaussian accuracy in central DP. Through experiments with both synthetic and real-world datasets,ExSubmechanism in the local model have been shown to reduce error by$40\%-60\%$compared to SOTA approaches. TheExSubin the shuffle model can further reduce over$85\%$error, and the online shuffle protocol reduces over$99.7\%$error.

In the federated learning scenario, geographically distributed clients collaboratively train a global model. Data heterogeneity among clients significantly results in inconsistent model updates, which evidently slow down model convergence. To alleviate this issue, many methods employ regularization terms to narrow the discrepancy between client-side local models and the server-side global model. However, these methods impose limitations on the ability to explore superior local models and ignore the valuable information in historical models. Besides, although the up-to-date representation method simultaneously concerns the global and historical local models, it suffers from unbearable computation cost. To accelerate convergence with low resource consumption, we innovatively propose a model regularization method named FedTrip, which is designed to restrict global-local divergence and decrease current-historical correlation for alleviating the negative effects derived from data heterogeneity. FedTrip helps the current local model to be close to the global model while keeping away from historical local models, which contributes to guaranteeing the consistency of local updates among clients and efficiently exploring superior local models with negligible additional computation cost on attaching operations. Empirically, we demonstrate the superiority of FedTrip via extensive evaluations. To achieve the target accuracy, FedTrip outperforms the state-of-the-art baselines in terms of significantly reducing the total overhead of client-server communication and local computation.

Federated Learning (FL) emerges as a distributed machine learning paradigm without end-user data transmission, effectively avoiding privacy leakage. Participating devices in FL are usually bandwidth-constrained, and the uplink is much slower than the downlink in wireless networks, which causes a severe uplink communication bottleneck. A prominent direction to alleviate this problem is federated dropout, which drops fractional weights of local models. However, existing federated dropout studies focus on random or ordered dropout and lack theoretical support, resulting in unguaranteed performance. In this paper, we propose Federated learning with Bayesian Inference-based Adaptive Dropout (FedBIAD), which regards weight rows of local models as probability distributions and adaptively drops partial weight rows based on importance indicators correlated with the trend of local training loss. By applying FedBIAD, each client adaptively selects a high-quality dropping pattern with accurate approximations and only transmits parameters of non-dropped weight rows to mitigate uplink costs while improving accuracy. Theoretical analysis demonstrates that the convergence rate of the average generalization error of FedBIAD is minimax optimal up to a squared logarithmic factor. Extensive experiments on image classification and next-word prediction show that compared with status quo approaches, FedBIAD provides 2× uplink reduction with an accuracy increase of up to 2.41% even on non-Independent and Identically Distributed (non-IID) data, which brings up to 72% decrease in training time.

As a popular privacy-preserving model training technique, Federated Learning (FL) enables multiple end-devices to collaboratively train Deep Neural Network (DNN) models without exposing local privately-owned data. According to the FL paradigm, resource-constrained end-devices in IoT should perform model training which is computation-intensive, whereas the edge server occupied with powerful computation capability only performs model aggregation. Due to the above unbalanced computation pattern, IoT-oriented FL is time-consuming and inefficient. In order to alleviate the computation burden of end-devices, recent countermeasures introduce the edge server to assist end-devices in model training. However, existing works neither efficiently address the computation heterogeneity across end-devices nor reduce the leakage risk of data privacy. To this end, we propose a Federated Synergy Learning (FedSyL) paradigm which innovatively strikes a balance between training efficiency and data leakage risk. We explore the complicated relationship between the local training latency and multi-dimensional training configurations, and design a uniform training latency prediction method by applying the polynomial quadratic regression analysis. Additionally, we design the optimal model offloading strategy with the consideration of resource limitation and computation heterogeneity of end-devices, so as to accurately assign capability=matched device-side sub-models for heterogeneous end-devices. We implement FedSyL on a real test-bed comprising multiple heterogeneous end-devices. Experimental results demonstrate the superiority of FedSyL on training efficiency and privacy protection.