Lei Yang

Machine learning models are trained on datasets where undesired features have spurious correlations with the target variable, resulting in the bias shifts of the model output to a different class. In federated learning (FL), a common problem is data distribution skew. Since each client has a different distribution, models trained on these local datasets are likely to learn different biases. This phenomenon will degrade the performance of the aggregated global model and then prevent the global model from learning actual intrinsic decisions. Consequently, addressing and mitigating these biases in FL is critical to the overall effectiveness of the global model. Most existing works are designed for centralized machine learning using data augmentation or regularization terms. Nevertheless, these methods may not generalize effectively when directly migrated to federated learning scenarios due to the lack of direct access to the external data of other clients. Recent work for debiasing in FL is based on data augmentation in image space, which generates new bias-conflicting images for further training, leading to high costs and limited ability to create sufficiently diverse and realistic samples. To tackle these challenges, we propose a novel feature-level augmentation approach, named FedInt, for debiasing in federated learning. The feature-level augmentation in latent space allows the model to capture and smooth out different directions of the decision boundary, making it possible to synthesize diverse bias-conflicting samples for each client with lower training costs. Experiments on three datasets demonstrate the effectiveness of our proposed method in terms of debiasing performance, and it outperforms traditional data augmentation techniques and various typical FL methods. Our code is available on an anonymous GitHub repository: https://anonymous.4open.science/r/FedInt-C710889456/ .

Knowledge Distillation (KD) is considered as an efficient way to replace the parameter averaging in federated learning, aiming to handle the clients with heterogeneous model architectures. Relying on the prepared distillation datasets across clients and the server, KD may encounter impractical difficulties in real-world implementations. Existing works explore the data-free KD in federated learning, which generates the distillation datasets on-site. However, the distillation datasets with global data distribution generated by these state-of-the-art schemes cannot be adapted to local non-IID data. In this article, we propose a new Personalized Data-Free Knowledge Distillation, namely PDKD, for federated learning under heterogeneous models and data. PDKD solves the problem of model drift caused by the inconsistent distribution of distillation datasets and the local data by generating personalized distillation datasets for each client while protecting client data privacy. In addition, we design a distillation dataset update scheme that maximizes the difference between teacher and client outputs on distillation datasets to accomplish deeper knowledge transfer. Furthermore, in order to accomplish the co-evolution of the teacher model and the clients’ model, PDKD incorporates a mutual distillation scheme. Numerous experiments show that PDKD significantly outperforms several state-of-the-art algorithms, with an 18% improvement in prediction accuracy and has a much lower communication cost than the compared algorithms.

Multimodal Large Language Models (MLLMs) extend foundation models to real-world applications by integrating inputs such as text and vision. However, their broad knowledge capacity raises growing concerns about privacy leakage, toxicity mitigation, and intellectual property violations. Machine Unlearning (MU) offers a practical solution by selectively forgetting targeted knowledge while preserving overall model utility. When applied to MLLMs, existing neuron-editing-based MU approaches face two fundamental challenges: (1) forgetting becomes inconsistent across modalities because existing point-wise attribution methods fail to capture the structured, layer-by-layer information flow that connects different modalities; and (2) general knowledge performance declines when sensitive neurons that also support important reasoning paths are pruned, as this disrupts the model's ability to generalize. To alleviate these limitations, we propose a multimodal influential neuron path editor (MIP-Editor) for MU. Our approach introduces modality-specific attribution scores to identify influential neuron paths responsible for encoding forget-set knowledge and applies influential-path-aware neuron-editing via representation misdirection. This strategy also enables effective and coordinated forgetting across modalities while preserving the model's general capabilities. Experimental results demonstrate that MIP-Editor achieves a superior unlearning performance on multimodal tasks, with a maximum forgetting rate of 87.75% and up to 54.26% improvement in general knowledge retention. On textual tasks, MIP-Editor achieves up to 80.65% forgetting and preserves 77.9% of general performance. Codes are available at https://github.com/PreckLi/MIP-Editor.

Edge AI applications enable edge devices to collaboratively learn a model via repeated model aggregations, aiming to utilize the distributed data on the devices for achieving high model accuracy. Existing methods either leverage a centralized server to directly aggregate the model updates from edge devices or need a central coordinator to group the edge devices for localized model aggregations. The centralized server (or coordinator) has a performance bottleneck and a high cost of collecting the global state needed for making the grouping decision in large-scale networks. In this paper, we propose an Autonomous Model Aggregation (AMA) method for large-scale decentralized learning on edge devices. Instead of needing a central coordinator to group the edge devices, AMA allows the edge devices to autonomously form groups using a highly efficient protocol, according to model functional similarity and historical grouping information. Moreover, AMA adopts a reinforcement learning approach to optimize the size of each group. Evaluation results on our self-developed edge computing testbed demonstrate that AMA outperforms the benchmark approaches by up to 20.71% in accuracy and reduced the convergence time by 75.58%.

Collaborative edge computing enables task execution on the computing resources of geo-distributed edge nodes. One of the key challenges in this field is to realize reliable task offloading by deciding whether to execute tasks locally or delegate them to neighboring nodes while ensuring task reliability. Achieving reliable task offloading is essential for preventing task failures and maintaining optimal system performance. Existing works commonly rely on task redundancy strategies, such as active or passive redundancy. However, these approaches lack adaptive redundancy mechanisms to respond to changes in the network environment, potentially resulting in resource wastage from excessive redundancy or task failures due to insufficient redundancy. In this work, we introduce a novel approach called Hybrid Redundancy for Task Offloading (HRTO) to optimize task latency and reliability. Specifically, HRTO utilizes deep reinforcement learning (DRL) to learn a task offloading policy that maximizes task success rates. With this policy, edge nodes dynamically adjust task redundancy levels based on real-time network load conditions and meanwhile assess whether the task instance is necessary for re-execution in case of task failure. Extensive experiments on real-world network topologies and a Kubernetes-based testbed evaluate the effectiveness of HRTO, showing a 14.6% increase in success rate over the benchmarks.

Time Sensitive Network (TSN) provides strict low latency and bounded jitter requirements for applications such as industrial systems, autonomous driving, etc. One of the important problems in TSN is to achieve high reliability and low latency by effectively routing and scheduling time-sensitive data flows. Existing work applies heuristic or integer programming to address flow routing and scheduling, yet often fail to achieve optimal solutions quickly. In this paper, we propose a new Reinforcement Learning (RL) based approach for routing and scheduling of redundant data flows, aiming to achieve load balancing on the network links as well as meeting the reliability and delay constraints. Our approach first leverages a simple heuristic algorithm to decide the redundant path candidate set, and then incorporates Proximal Policy Optimization (PPO) method to choose the most suitable multi-routing flows from the candidates, which can be aware of the network status dynamically to reduce the load on the bottleneck link of the network. On this basis, we further retrain the RL model by fine-tuning to adapt to the online environment. The simulation results show that our proposed solution outperforms the benchmark algorithms in terms of the degree of network balance by 38.7% in offline network environments and in terms of average delay by 14.0% in online network environments.

Federated Learning (FL) enables multiple data owners to jointly train a machine learning model, which can improve joint environmental cognitive capability without disclosing their private local data. In FL systems, local clients are often heterogeneous, characterized by different data distribution forms and varying levels of noise patterns. This poses challenges to the convergence and test accuracy of the learned global model. Existing client selection methods failed to effectively address the data heterogeneity problem while considering the impacts of correlations between clients. In this work, we propose an efficient and adaptive client selection scheme with the Federated Proximal Policy Optimization (FedPPO). It utilizes a noisy client filtering method that leverages the principal component analysis and hierarchical clustering to reduce the action search space and accelerate the convergence of Deep Reinforcement Learning (DRL). It then evaluates each client’s contribution by jointly using the accuracy of global model and local models and take the contribution score as the reward. Finally, it leverages a DRL algorithm, PPO, to adaptively select clients with high contributions to participate in training. Theoretical analysis shows that it is guaranteed to converge in an efficient manner. Experimental results on four real-world datasets show that FedPPO significantly outperforms existing approaches of constructing global models with faster convergence speed and higher test accuracy.

The implementation of intelligent driving technology is inseparable from reliable wireless communication and efficient cache scheduling in Vehicle Ad-hoc Networks (VANETs). The purpose of cache scheduling in VANETs is to provide efficient and fast data services to vehicles. In response to difficult to handle wireless communication interference and high requirements for communication delay and energy consumption optimization in vehicle-road cooperation, we propose a new communication model of “Retrieving-Cast” to simplify the cache scheduling process oriented by data requirements. In V2I mode, consider the uplink of all roadside units(RSUs) retrieving data from vehicles, and prioritize data retrieving based on factors such as the number, density, and urgency of vehicles within the communication coverage range of RSUs. We hope that as many high-priority RSUs as possible to complete cached data retrieving as soon as possible under wireless communication interference. To overcome the limitations of traditional heuristic algorithms with low scheduling efficiency and deep learning methods requiring a large amount of training data, we propose a double-$\varepsilon$ online reinforcement learning method, which efficiently solves the cache retrieval scheduling problem. Through experiments and analysis, our method achieved better scheduling results compared to traditional approximation algorithms and single $\varepsilon$-greedy Deep Q-learning methods. The designed data transformation method also addresses the issue of requiring a large amount of training data, enabling rapid computation of scheduling requests in the network at an extremely low training cost, meeting the low-latency computation requirements in VANETs.

Distributed machine learning on edges is widely used in intelligent transportation, smart home, industrial manufacturing, and underground pipe network monitoring to achieve low latency and real time data processing and prediction. However, the presence of a large number of sensing and edge devices with limited computing, storage, and communication capabilities prevents the deployment of huge machine learning models and hinders its application. At the same time, although distributed machine learning on edges forms an emerging and rapidly growing research area, there has not been a systematic survey on this topic. The article begins by detailing the challenges of distributed machine learning in edge environments, such as limited node resources, data heterogeneity, privacy, security issues, and summarizes common metrics for model optimization. We then present a detailed analysis of parallelism patterns, distributed architectures, and model communication and aggregation schemes in edge computing. we subsequently present a comprehensive classification and intensive description of node resource-constrained processing, heterogeneous data processing, attacks and protection of privacy. The article ends by summarizing the applications of distributed machine learning in edge computing and presenting problems and challenges for further research.