Online model training is pivotal for enabling multiuser semantic communication systems to adapt to dynamic channel conditions. However, conventional frameworks suffer from prohibitive communication overhead and vulnerabilities to privacy attacks, hindering practical deployment. This paper proposes semantic information mixup (SIMix), a secure and efficient training framework that integrates Over-the-Air Mixup (OAM) with label-aware user grouping to jointly optimize spectral efficiency and semantic security. The OAM mixes semantic features of multiple users via wireless channels, inherently obfuscating sensitive data while reducing communication overhead. A closed-form Tx-Rx scaling optimization minimizes the mean square error (MSE) of over-the-air computation under channel noise, ensuring stable convergence in low-SNR regimes. Furthermore, an extended max-clique algorithm dynamically partitions users into groups with minimal intra-label similarity, reducing model inversion attack success rates. Experiments on CIFAR-10 and Tiny ImageNet demonstrate that the proposed approach is superior in terms of communication efficiency and security, reducing communication overhead by up to 25% and attaining 17.58 dB PSNR (20.98 dB reduction) under inversion attack and reducing 13.44% attack success rate under label inference attack, while achieving comparable transmission accuracy.

Vertical Federated Learning (VFL) enables collaborative multi-agent model training with distributed feature spaces. Compared to horizontal federated learning (HFL), poisoning attacks in VFL confront the challenges of partial/incomplete data and model information, i.e., the adversary only has its local input feature space and model structure without knowing the ground-truth labels or server-side model. Existing attacks are inefficient in terms of stability and scalability under different learning configurations. In this paper, we design the poisoning framework via embedding manipulation in the model-partitioning structure of VFL, and propose two efficient poisoning attacks, i.e., layer embedding manipulation (LMP) and perturbation embedding manipulation (PMP). Both LMP and PMP meticulously transform the benign embeddings into the poisoned ones to dominate the model prediction. In particular, LMP blends a malicious trigger layer to poison the embedding. PMP exploits the gradients to generate a universal perturbation, which can be injected into the benign embeddings for poisoning. We conduct extensive experiments for attack performance evaluation. PMP and LMP achieve an average attack success rate of more than 0.9, and keep stability and robustness under different learning configurations. We further conduct six different defense methods to evaluate the proposed attacks. The proposed attacks are still shown to succeed against these defenses, necessitating the advanced defenses for VFL in future work.

Input diversity is an effective technique for crafting transferable adversarial examples that can deceive unknown AI models. Existing input-diversity-based methods typically use single input transformation, limiting targeted transferability and defense robustness. Combining different transformation types is challenging, as keeping increasing types would degrade semantic information and targeted transferability. This paper proposes a quality-awaretransformationcombinationattack (TCA) that selects high-quality transformation combinations. The quality-aware selection enables expansion of transformation types, enhances input diversity, and hence improves targeted transferability and defense robustness. We first design a quality-evaluation framework to quantify the effectiveness of transformation combinations, which jointly considers convergence, transferability, and robustness. Only a small group (up to 10) of images are required for computation-efficient quality evaluation. Experiments validate TCA's superiority over state-of-the-art baselines in adversarial transferability and robustness. When defenses are secured, the average targeted success rate of TCA with four transformation types (i.e., TCA-t4) outperforms the best baseline by 26%$\sim$42% on ImageNet.

Distributed learning is a promising paradigm for future UAV (unmanned aerial vehicle) networks networks and other emerging autonomous unmanned systems. Such distributed learning framework can suit the intrisic decentralized topology of UAV networks, where the UAVs can collaborate to train a global AI model by only exchanging the model parmeters via its peer-to-peer (i.e, inter-UAV) links in a distributed manner. However, with the ever-increasing AI model sizes, the challenges arise from the significant communication overhead for exchanging massive model weights via inter-UAV links in an ad-hoc manner: Previous communication-efficient techniques are mainly designed for conventional federated learning and not easily extendable to the decentralized counterpart. We propose selective link orchestration to minimize communication overhead while ensuring convergence of distributed learning, and prove that the convergence constraint is equivalent to the connectivity of the selected sub-graph. As such, we can reformulate the problem as a link selection problem in graph theory and develop a distributed optimization algorithm based on the modification of the Gallager, Humblet, and Spira’s algorithm. Experimental results on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate up to a $90 \%$ reduction in communication overhead without compromising model accuracy.

Federated learning is a popular distributed machine learning paradigm that enables collaborative model training at multiple entities via exchanging intermediate learning results. Security and communication efficiency are crucial for successful applications of federated learning in various privacy-sensitive services. However, existing work focused on gradient defense and communication efficiency separately, and also incurred additional computation, signaling, and accuracy overhead. A lightweight (in terms of time-complexity and signaling) technique that simultaneously achieves security and communication efficiency is critical for massive resource-constrained devices (e.g., Internet-of-Things generating the data), but has yet to be established. This paper proposes a secure and efficient federated learning framework with provable communication-accuracy-security performance guarantees. A low-complexity and signaling-free stochastic quantization module is added at the client side that quantizes the original local gradients to discrete values for communication-efficient global aggregation. The stochastic quantization module is shown to be interpreted as triangular or Gaussian-multiply-triangular noises under uniform or Gaussian distributions of local gradients, hence protecting data privacy. We prove that the proposed framework exhibits an {O(log21/δ),O(δ2),O(1/δ)}-tradeoff between the communication overhead, model accuracy, and data protection, where δ is an adjustable quantization interval. Experimental results validate the tradeoff and the superiority of the proposed stochastic quantization technique in terms of communication efficiency (only 14.1% of differential privacy and 0.2% of homomorphic encryption) and computation complexity (similar to differential privacy and only 0.03% of homomorphic encryption). Under the same data protection performance, the proposed approach also outperforms (in terms of accuracy) differential privacy in all the 9 comparison settings on CIFAR10 dataset.

Resource provisioning aims to efficiently allocate the infrastructure provider's communication, computation and caching resources to provide reliable (e.g., 99.999%) and responsive actions for massive delay-critical services. However, the challenges arise from the partial knowledge of uncertain traffic arrivals (even the distributions of estimation error may not be available) and computation complexity for massive numbers of emerging services. This paper aims to design efficient and robust resource provisioning for massive delay-critical services based on only partial knowledge on traffic uncertainty. We formulate the problem of robust resource provisioning to minimize the system cost with only partial (i.e., the first/second momentum) information of traffic estimation errors. We derive the robust approximation of the reliability constraint using the Bernstein approximation, which is proved to guarantee service reliability given the partial traffic knowledge. The problem is reformulated and solved via Lagrangian duality to obtain the closed-form expression for low-complexity solutions. Experimental results on both the simulation-based and trace-based datasets validate that the proposed approach can guarantee up to 99.999% service reliability with reduced time complexity.

Split learning is an emerging distributed machine learning framework for enabling edge intelligence, especially for training sophisticated AI models at resource-constrained Internet-of-Things (IoT) devices. In split learning, the full AI model is partitioned to the client-side (e.g., input/privacy-sensitive layers) and server-side (e.g., computation-intensive layers) portions to be trained collaboratively at the devices and edge server. Only intermediate data at the split layer are exchanged during the training process for data privacy. However, the intermediate data may still cause security concerns to reconstruct the raw data from the partial gradients. This paper proposes the masking-enabled data protection approach for split learning without compromising the model accuracy. The devices are designed to perturb the reported results via masks, and the adversary can only retrieve the global information of all the devices (instead of individual devices). We mathematically prove that the masking-enabled perturbation mechanism would not compromise the learning accuracy. Experimental results validate the effectiveness of the proposed approach in terms of successful data protection and up to 10% model accuracy gain, compared to vanilla split learning and differential privacy.

The Internet of Vehicles (IoV) is a promising paradigm for autonomous driving, where the sensing data from the onboard sensors can be disseminated and processed cooperatively via vehicle-to-vehicle links. Autonomous vehicles can share their local views for cooperative, reliable, and robust driving decisions. However, the limited wireless resources may become the bottleneck with the increasing number of vehicles. The technical challenges also arise from the decentralized control, the spatial couplings of decisions, and the complexity of combinatorial optimization. This paper proposes a novel fully distributed graph-based approach to jointly optimize multicast link establishment with data dissemination and processing decisions by only exchanging partial information among neighboring vehicles. The mixed-integer programming problem aims to maximize system energy efficiency while achieving maximum data throughput. We prove that maximizing data processing throughput is submodular optimization to find the local optimum efficiently. The optimization of data dissemination and processing is reformulated to a minimum-cost maximum-flow problem in a three-layer graph, and efficiently solved by exploiting the graphical interdependence. Both simulation-generated and trace-based datasets are evaluated to validate the effectiveness of the proposed approach in terms of data throughput and energy efficiency.

Cooperative caching can unify storage across edge clouds and provide efficient delivery of popular contents under effective content placement. However, the placement and delivery are non-trivial in cooperative caching due to the decentralized property of edge clouds, as well as the temporal and spatial correlation of the placement. We propose a new distributed online learning approach to jointly optimize content placement and delivery without the a-priori knowledge on file popularity and link availability. Content placement and delivery can be asymptotically optimized in real-time by running distributed online learning at individual edge servers by exploiting stochastic gradient descent (SGD). The proposed approach can allow operations at different timescales by integrating mini-batch learning for farsighted content placement. The optimality loss, stemming from the different timescales, can asymptotically reduce, as the SGD stepsize declines. Simulations confirm that the proposed approach outperforms existing techniques in terms of cache hit ratio and cost effectiveness. Insights are shed on the optimal placement of popular contents.

This paper aims to produce an effective online scheduling technique, where a base station (BS) schedules the transmissions of energy harvesting-powered Internet-of-Things (IoT) devices only based on the (differently outdated) in-band reports of the devices on their states. We establish a new primal-dual learning framework, which learns online the optimal proactive schedules to maximize the time-average throughput of all the devices. Batch gradient descent is designed to enable stochastic gradient descent (SGD)-based dual learning to learn the network dynamics from the outdated reports. Replay memory is deployed to allow online convex optimization (OCO)-based primal learning to predict channel conditions and prevent over-fitting. We also decentralize the online learning between the BS and devices, and speed up learning by leveraging the instantaneous knowledge of the devices on their states. We prove that the proposed framework asymptotically converges to the global optimum, and the impact of the outdated knowledge of the BS diminishes. Simulation results confirm that the proposed approach can increasingly outperform state of the art, as the number of devices grows.

Lyapunov optimization has shown to be effective for online optimization of fog computing, asymptotically approaching the optimality only achievable offline. However, it is not directly applicable to the Internet of Things, as inexpensive sensors have small buffers and cannot generate sufficient backlogs to activate the optimization. This article proposes an enabling technique for the Lyapunov optimization to operate under finite buffers without loss of asymptotic optimality. This is achieved by optimizing the biases (namely, “virtual placeholders”) of the buffers to create sufficient backlogs. The optimization of the placeholders is proved to be a new three-layer shortest path problem and solved in a distributed manner by extending the Bellman-Ford algorithm. The sizes of the virtual placeholders decline fastest along the shortest paths from the sensors to the data center, thereby preventing unnecessary detours and reducing end-to-end delays. Corroborated by simulations, the proposed approach is able to operate under the conditions the direct application of the Lyapunov optimization fails, and significantly increase the throughput and reduce the delays in other cases.

Processing data around the point of capture, fog computing can support computationally demanding Internet-of-Things (IoT) services. Distributed online optimization is important given the size of IoT, but challenging due to time variations of random traffic and nonuniform connectivity (or cardinality) of edge servers and IoT devices. This paper presents a distributed online learning approach to asymptotically minimizing the time-average cost of fog computing in the absence of the a-priori knowledge on traffic randomness, for light-weight, and delay-tolerant application scenarios. Stochastic gradient descent is exploited to decouple the optimizations between time slots. A graph matching problem is then formulated for every time slot by decoupling and unifying the nonuniform cardinalities, and solved in a distributed manner by developing a new linear (1/2)-approximation method. We prove that the optimality loss resulting from the distributed approximate graph matching method can be compensated and diminish by increasing the learning time. Corroborated by simulations, the proposed distributed online learning is asymptotically optimal and superior to the state of the art in terms of throughput and energy efficiency.

To enable machine learning at the edge of wireless networks (such as edge cloud), close to mobile users, is critical for future wireless networks, but challenging since the lower layers in edge cloud are substantially different from existing machine learning configurations in the cloud. In such geo-distributed computing environment, streaming data need to be evenly and cost-efficiently partitioned for different workers to produce an unbiased learning model with reduced parameter synchronization frequency. This paper presents a new online approach to optimally partitioning streaming data under time-varying network conditions. A new measure is proposed to quantify the evenness of data partitioning and restrain the optimization of data admission, partitioning, and processing. Stochastic gradient descent is applied to learn the optimal decisions online and asymptotically maximize the time-average utility of data partitioning. A new protocol is designed to further reduce the measurements of link costs, while preserving the asymptotic optimality, data evenness, and stability of the platform. Simulation results show that the proposed approach is superior to the state of the art in terms of throughput and cost efficiency, while only 24% of the links need to be measured to achieve the asymptotic optimality.

Cooperative caching can unify network storage to improve efficiency, but the effective placement and search of contents are challenging especially in distributed edge clouds with neither a-priori knowledge on content requests nor instantaneous global view. This paper establishes a new profitable cooperative region for every content request admitted at an edge server, within which the content, if cached, can be retrieved with guaranteed profit against a direct retrieval from the network backbone. This narrows down the search for the content. The caching density of the content can also be significantly reduced, e.g., to a cached copy per region. The regions are based on a novel distributed framework which allows individual servers to spontaneously admit/dispatch requests and deliver/forward contents, while asymptotically maximizing the time-average profit of caching. The cooperative region for content is erected at individual servers by comparing the upper and lower bounds for the backlogs of unsatisfied requests of the content. Simulations show the substantially improved profit of the proposed approach over existing solutions. The regions can help automate the placement of contents with reduced density and improved efficiency.

Fog computing enables resource-limited network devices to help each other with computationally demanding tasks, but has yet to be implemented in large scales due to sophisticated control and network inhomogeneity. This paper presents a new fully distributed online optimization to asymptotically minimize the time-average cost of fog computing, where tasks are selected to be offloaded and processed independently between different links and devices by measuring their cost effectiveness at each time slot. A key contribution is that we optimize the cost-effectiveness measures which achieve the asymptotic optimality over infinite time. Another contribution is that we optimize placeholders at the devices; which create collaborative computing regions of tasks in the vicinity of the point of capture, prevent tasks being offloaded beyond, preserve the asymptotic optimality and reduce delay. This is achieved in a distributed fashion by discovering the optimal substructure of the placeholders. Simulations show that the average size of collaborative regions is only 3.2 out of total 500 servers, and the system income increases by 43% as compared with existing techniques.

Decentralized orchestration of the control plane is critical to the scalability and reliability of software-defined network (SDN). However, existing orchestrations of SDN are either one-off or centralized, and would be inefficient the presence of temporal and spatial variations in traffic requests. In this paper, a fully distributed orchestration is proposed to minimize the time-average cost of SDN, adapting to the variations. This is achieved by stochastically optimizing the on-demand activation of controllers, adaptive association of controllers and switches, and real-time request processing and dispatching. The proposed approach is able to operate at multiple timescales for activation and association of controllers, and request processing and dispatching, thereby alleviating potential service interruptions caused by orchestration. A new analytic framework is developed to confirm the asymptotic optimality of the proposed approach in the presence of non-negligible signaling delays between controllers. Corroborated from extensive simulations, the proposed approach can save up to 73% the time-average operational cost of SDN, as compared to the existing static orchestration.