With the widespread popularity of Large Language Models (LLMs), the mixture of experts (MoE) has not only emerged as a key enabler for scaling up model capacity by significantly reducing computational demands, but also for giving rise to edge-computing empowered distributed LLMs with better prices, low latency, and regional privacy. Nonetheless, besides constraints in computational capability, edge-based LLM deployments also face challenges such as unreliable environments due to the limited security of edge devices. In this paper, we propose REMIS, an inference task scheduling scheme designed to enable MoE-based LLM services under untrustworthy computation conditions. Specifically, after an LLM is properly partitioned into shards and deployed across edge devices, REMIS dynamically schedules and activates experts on devices with lower loads and higher reliability. This strategy is effectively achieved through a deep reinforcement learning procedure that optimizes both servicing latency and inference credibility. Unlike most existing MoE-based schemes with fixed Top-K routing, REMIS operates in a novel plug-in manner, intelligently selecting experts to improve task adaptability. Numerical evaluations under various untrustworthy setups validate the superiority of our proposed scheme in both servicing latency and inference credibility.

Presents corrections to the paper, (Corrections to “Giant Could Be Tiny: Efficient Inference of Giant Models on Resource-Constrained UAVs”).

Within the Internet of Things (IoT) landscape, Encoder as a Service (EaaS) is a cloud-based service for many AI empowered scenarios, such as autopilot and face-scan payment, enabling IoT devices to keep a light classifier model at local while remotely accessing powerful encoding models. However, adversarial examples like an image with imperceptible tiny perturbations that can lead to incorrect classifying results, make it challenging to achieve a robust EaaS-based classifier. Certified radius (R) emerges as a measure against such imperceptible tiny perturbations, and guarantees the trustworthiness of any input with perturbations smaller than R. Despite offering R related services, the substantial computational overhead from traditional methods, stemming from the extensive searches of R for each request, poses a barrier to their practical deployment. In this article, we identify the repetitive computations in the EaaS framework due to the fixed encoding model and propose a novel search cache scheme to speed up the R-related computation. Therefore, we propose large-scale efficient robust EaaS (ScaleRES) to address different R-related services: First, ScaleRES strategically stores previous computation results of R, to enable the reuse and refining of R across users. Then, ScaleRES obtains the average certified radius (ACR) efficiently with selective cached R. Finally, ScaleRES performs R filtering to enhance adversarial training for a robust EaaS-based classifier. Comprehensive evaluations demonstrate that in three distinct computations, ScaleRES offers significant savings in computational overhead compared to the conventional approach—70% for R computation as the number of clients increases, 35% for ACR of the entire test set, and 40% for adversarial training with robustness comparable to other works.

With the rapid expansion of IoT and the advent of the 6G era, ensuring efficient and secure communication for the edge large language model (LLM) servicing has become a critical priority. However, malicious relay nodes injecting delays pose a significant threat to network performance and security. This paper proposes a novel framework combining Stackelberg Game theory with Multi-Agent Reinforcement Learning (MARL) to mitigate communication delay injection attacks in edge networks. By modeling the interaction between edge devices and malicious nodes as a Stackelberg Game, where edge devices act as followers optimizing relay strategies and malicious nodes as leaders seeking disruption, our approach enables edge devices to cooperatively identify and avoid malicious nodes. The distributed MARL framework allows each edge device to learn their optimal strategies independently, enhancing network resilience. Experimental results demonstrate the effectiveness of our scheme in effectively reducing the impact of malicious nodes and improving the deployment capability of edge computing services.

Graph neural networks (GNNs) have shown great success in IoT applications, but many IoT scenarios further involve evolving graph data over time. Streaming graph learning (SGL) tackles the issue by updating GNN models continuously, incorporating significant historical node data for tasks like node classification. However, existing research on SGL often neglects memory limitations during historical node selection, and efficient distributed training is challenging due to the coupling of node selection, placement, and parallelization. This article proposes an efficient distributed SGL system that optimizes node selection and storage in multi-GPU environments, while reducing communication overhead to accelerate training. Extensive evaluations demonstrate the effectiveness of the approach.

Binary malware summarization aims to automatically generate human-readable descriptions of malware behaviors from executable files, facilitating tasks like malware cracking and detection. Previous methods based on Large Language Models (LLMs) have shown great promise. However, they still face significant issues, including poor usability, inaccurate explanations, and incomplete summaries, primarily due to the obscure pseudocode structure and the lack of malware training summaries. Further, calling relationships between functions, which involve the rich interactions within a binary malware, remain largely underexplored. To this end, we propose MALSIGHT, a novel code summarization framework that can iteratively generate descriptions of binary malware by exploring malicious source code and benign pseudocode. Specifically, we construct the first malware summary dataset, MalS and MalP, using an LLM and manually refine this dataset with human effort. At the training stage, we tune our proposed MalT5, a novel LLM-based code model, on the MalS and benign pseudocode datasets. Then, at the test stage, we iteratively feed the pseudocode functions into MalT5 to obtain the summary. Such a procedure facilitates the understanding of pseudocode structure and captures the intricate interactions between functions, thereby benefiting summaries’ usability, accuracy, and completeness. Additionally, we propose a novel evaluation benchmark, BLEURT-sum, to measure the quality of summaries. Experiments on three datasets show the effectiveness of the proposed MALSIGHT. Notably, our proposed MalT5, with only 0.77B parameters, delivers comparable performance to much larger Code-Llama.

Giant models, characterized by their billions or even trillions of parameters, has demonstrated unprecedented capabilities in handling complex tasks on Artificial intelligence (AI)-driven UAVs, such as disaster relief, aerial navigation, and manipulation. However, there is an open challenge about the mismatching between the massive computation and memory requirements of giant models and the limited resources on UAVs. Existing works either pose privacy concerns with offloading methods or compromise model accuracy with various model compression techniques. In this paper, we fill the gap by exploiting the Mixture-of-Expert (MoE) model architecture that decouples giant models into multiple tiny experts, so that UAVs can dynamically load a few experts that best match their current input. We consider a general scenario of several edge servers feeding experts to multiple UVAs and formulate a core problem of expert selection and UAV-edge association. Due to the high complexity of this problem, we propose a solution, termed GESolver, based on graph learning, which automatically solves the problem by learning the complicated interaction between edge servers, UAVs, as well as their required experts. We evaluate our proposed method with three popular MoE-based models under various problem settings. The experiments demonstrate that our proposed method can significantly outperform other baselines.

Due to the non-Gaussian distribution of erratic noise, conventional Gaussian denoising methods often encounter substantial challenges and pressures when suppressing this kind of noise. To overcome this challenge, several state-of-the-art (SOTA) schemes, for instance, robust low-rank approximation (LRA) and deep learning (DL) methods, have been designed and achieved promising results in the treatment of erratic noise. However, these SOTA denoising methods focus mainly on matrix-based modeling representations and fail to fully reflect the correlations associated with erratic noise and valid signals in the spatial dimension and thus may display suboptimal performance. As an alternative, a robust tensor DL (RTDL) denoising method for unsupervised 3-D seismic erratic noise suppression that involves the use of a reasonable combination of tensor sparse representation (SR) and a tensor neural network (tNN) is proposed in this study. The key to RTDL is to introduce a robust tensor sparse norm for erratic noise to exhaustively exploit the spatial tubular sparse distribution properties in 3-D space; notably, adding a tensor sparse norm to the tNN model yields a new data-driven model with 3-D erratic noise reduction capabilities. To find the optimized parameters of the new model, an efficiency tensor optimization method is established on the basis of alternating minimization (Alt), the aim of which is to alternately solve two subproblems involving tensor SR and a tNN. This paper presents well-designed experiments and satisfactory results compared with those of SOTA methods based on both synthetic and real field datasets.

It is vital to closely track the operation statuses of network-internal links. Accurate knowledge of the operation statuses of network-internal links is vital for the management of many networks like the Internet, the satellite communication network, etc. Network boolean tomography can identify congested links just using end-to-end path status observations, and is able to work efficiently even without any available cooperation of internal nodes. Nevertheless, it heavily assumes that all the path status observations collected are true while some Byzantine attacks, e.g., the label flip attacks, could violate this assumption. In this paper, we present a performance evaluation of network boolean tomography under Byzantine attacks. Our results against various attacking rates, locations, and scales all show that Byzantine attacks could cause a significant performance degradation of network boolean tomography, suggesting a pressing need of developing the detection and countermeasure techniques.

Data privacy and security issues are preventing a lot of potential on-cloud machine learning as services from happening. In the recent past, secure multi-party computation (MPC) has been used to achieve the secure neural network predictions, guaranteeing the privacy of data. However, the cost of the existing two-party solutions is expensive and they are impractical in real-world setting.

The core challenge of seismic data interpolation is how to capture latent spatial-temporal relationships between unknown and known traces in 3-D space. The prevailing tensor-based interpolation schemes seek a globally low-rank approximation to mine the high-dimensional relationships hidden in 3-D seismic data. However, when the low-rank assumption is violated for data involving complex geological structures, the existing interpolation schemes fail to precisely capture the trace relationships, which may influence the interpolation results. As an alternative, this article presents a basic deep tensor autoencoder (DTAE) and two variants to implicitly learn a data-driven, nonlinear, and high-dimensional mapping to explore the complicated relationship among traces without the need for any underlying assumption. Then, tensor backpropagation (TBP), which can be essentially viewed as a tensor version of traditional backpropagation (BP), is introduced to solve for the new model parameters. For ease of implementation, a mathematical relationship between tensor and matrix autoencoders is constructed by taking advantage of the properties of a tensor–tensor product. Based on the derived relationship, the DTAE weight parameters are inferred by applying a matrix autoencoder to each frontal slice in the discrete cosine transform (DCT) domain, and this process is further summarized into a general theoretical and practical framework. Finally, the performance benefits of the proposed DTAE-based method are demonstrated in experiments with both synthetic and real field seismic data.

Human behavior recognition which is the indispensable technology for Artificial Intelligence(AI) application like smart home and other practical applications, is very challenging as the optimal recognition generally is required to be non-invasive and easy to deploy. An increasing interest has been paid on the human behavior recognition with off-the-shelf Wi-Fi devices. However, most of existing works just limit their focus on the small-scale scene while human behavior recognition will be quite different in large areas for a larger number of antennas and correspondingly a more complex antenna layout. For example, if we want to build a complete behavior awareness system using the distributed Wi-Fi equipment of the entire building, though collecting and using all antennas’ data is feasible maybe, the overhead concerns of computing and bandwidth resources, and the operation complexity will be hard to lessen in practice. In this paper, we first present analyses of the signal performances between different antenna pairs. Then closely following these analyses, we propose a novel scheme for the large-area human behavior recognition. Finally, we conduct extensive confirmatory experiments to verify the validity of our proposed scheme.

Efficient urban traffic monitoring is a key enabler for intelligent planning and management of modern cities. Network tomography can monitor the urban traffic with a comparably small number of traffic detectors like cameras, and has become an appealing technique for urban traffic management. However, previous work on network tomography based traffic monitoring focuses primarily on developing estimators using the given end-to-end travel time measurements, while the design of data collection for efficiently distributed collecting and processing the raw monitoring videos to such measurements is often neglected. We fill this gap by exploring the vision of edge intelligence for optimal urban traffic monitoring, and tackle the following two problems in regard of limited telecommunications resources: 1) when the total number of monitoring videos that are successfully processed into the end-to-end travel time measurements is pre-bounded, we employ a Fisher Information Matrix (FIM) to help determine the best quota scheme for the monitoring videos that each traffic detector need to generate and 2) when the centralised processing of monitoring videos alone is insufficient, we make use of the computation capabilities from these edge devices, i.e., traffic detectors, and employ a multi-agent reinforcement learning approach to help them conduct intelligent computation offloading individually. Extensive simulations demonstrate that our proposed scheme effectively reduces the estimation error of network tomography compared to common approaches with either uniform or random strategy.

The accurate identification of congested links is crucial for network performance monitoring. Network boolean tomography uses end-to-end path measurements to identify congested links, and appears as a significant alternative when direct link monitoring is not available. However, most of existing tomographic methods assume no correlations between links, i.e., the congestion of one link is assumed to be independent from the congestion of any others, hindering their applications in practice because links could become correlated during a joint optimization procedure of many network operations like traffic routing and balancing. In this paper, we study practical network boolean tomography without such an assumption. We elaborate on the ill-posed nature of network boolean tomography to highlight the significance of integrating link correlations, and model the congested link identification from end-to-end congestion observations of paths as a problem of Maximum A-Posteriori (MAP) estimation. To avoid the explicit acquisition of any priori knowledge of link correlations, we then propose a learning-based algorithm with Long Short-term Memory (LSTM), a special recurrent neural network that is good at learning statistical dependencies of sequence elements from historical data. Numerical results over real network topologies validate our learning-based network boolean tomography.

Mobile Edge Computing (MEC) enables mobile users to offload their computation loads to nearby edge servers, and is seen to be integrated in the 5G architecture to support a variety of low-latency applications and services. However, an edge server might soon be overloaded when its computation resources are heavily requested, and would then fail to process all of its received computation loads in time. Unlike most of existing schemes that ingeniously instruct the overloaded edge server to transfer computation loads to the remote cloud, we make use of the spare computation resources from other local edge servers by specially taking the risk of network link failures into account. We measure such link failure risks with the financial risk management metric of Conditional Value-at-Risk (CVaR), and well constrain it to the offloading decisions using a Minimum Cost Flow (MCF) problem formulation. Numerical results validate the enhancement of the MEC service's availability by our risk-aware offloading scheme.

Network congestion will result in significant performance degradation or even failures of many bandwidth-hungry Internet of Things (IoT) applications. Accurate and efficient congested link identification has become a foundational issue to IoT applications like self-driving cars, digital health, smart city, and so on. However, directly monitoring the massive number of interior links often introduces high operation cost or even is infeasible in practice, giving rise to indirect monitoring techniques like network Boolean tomography. Nevertheless, in many networks, the number of their interior links is larger than their end-to-end paths, making it very challenging for network Boolean tomography to find a determined solution. To resolve this issue, most of current methods try to utilize some prerequisites, such as the link congestion probabilities. While these probabilities might be hard or even unable to be obtained accurately in dynamical networks, limiting the practical deployment. In this paper, we are motivated to design a framework of congested link identification without any prerequisite or assumption. We first novelly model the congested link identification procedures as a Markov decision processes (MDPs), and then employ a reinforcement learning technology, i.e., Q-learning, to solve this MDP. The simulation results show that our proposed scheme can autonomously and efficiently explore the unknown network environment, and is able to achieve better adaptivity and correctness, without any prior knowledge comparing to existing methods.

Multipath routing, which is increasingly common in today's Internet, can lead two end-hosts to own multiple routing paths. This will impose adverse effects on most of the current network measurement methods, as they generally need to assume a single active path between any pair of end-hosts at any given time. In this paper, we attempt to identify whether multipath routing exists between two end-hosts using end-to-end packet order. We show theoretically that the probability of observing no out-of-order delivery among a strip of packets is always lower when the two end-hosts get multiple instead of a single path to forward them. Based on such investigation, we design a probe and use it to propose an end-to-end measurement scheme that can simultaneously identify multipath routing and its type (e.g., flow-based or packet-based). Both experiment and simulation results validate the effectiveness of our proposed scheme.

Congestion links can not only introduce great packet losses, but also cause significant delay flutters to paths that traverse them. However, most of current approaches just try to identify congestion links that meet end-to-end loss observations. What's worse, most of them also take no consideration of multipath routing, while which will introduce more than a single routing path between two end-hosts and can make a single-source network own a non-tree topology instead of the tree one. In this paper, we employ both end-to-end loss and delay observations to identify congestion links in a single-source network where multipath routing is enabled. We first prove that under certain topology conditions, the link delay variances in such non-tree topology can be inferred solely from end-to-end delay measurements. Then, we propose an algorithm to identify as congested a set of links, which can not only account for end-to-end path losses but also demonstrate great delay variances at the meantime. Simulation results validate the desirable performance of our proposed scheme.

Geographic routing in wireless sensor networks suffers from the local minimum problem when data packets encounter the concave boundary of a hole (void area), i.e., no neighbor node are closer to the destination than the current node. In this paper, we propose a hole plastic scheme which adaptively fill the concave area of a hole with potential stuck nodes according to 1-hop information of neighbors, to relieve the local minimum problem faced by geographic routing. The basic idea is To mark the nodes located in the concave area of the hole as potential stuck nodes which do not participate in data delivery unless a source/destination is located on the concave area of the hole. Once the hole plastic process is achieved, subsequently arriving data flows will be prevented from entering the concave area of the hole by potential stuck nodes. The proposed hole plastic scheme is achieved in a local self-organized manner, i.e., potential stuck node marking is based on the information of 1-hop neighbors, and traditional multi-hop cooperation methods such as hole detection, hole boundary tracing, hole modeling are not required.