By reducing the size of transmitted data between device-side and edge-side machine learning model parts, intermediate activation (IA) compression can alleviate communication overhead, lower latency, and conserve energy, thus enhancing the model partitioning in edge-device collaborative inference scenarios. However, existing studies lack refined resource allocation and fail to jointly optimize edge-device communication and computing resources, especially in terms of IA compression rates, leading to sub-optimal inference performance. To this end, we propose an adaptive model partitioning, intermediate activation compression, and resource allocation scheme for efficient edge-device collaborative inference. We jointly optimize the model partitioning point selection, IA compression rates control, computing resource allocation for both edge and devices, and device transmission power allocation. Our goal is to minimize the weighted sum of inference accuracy loss, inference latency, and device energy consumption. To solve the high-complexity optimization problem efficiently, we design a DRL-based algorithm, which decouples the problem into sub-problems firstly, and then employs an SD3 (Softmax Deep Double Deterministic Policy Gradients)-based DRL method to solve the partitioning point and IA compression rates sub-problem, and utilizes various numerical methods to solve the sub-problems of local and edge computing resource allocation and transmission power control. Extensive comparative simulations with four different schemes under different environmental parameters demonstrate the superiority and robustness of our approach.

Multi-modal learning improves model robustness and accuracy by integrating complementary information from multiple modalities, addressing the limitations of uni-modal approaches in handling complex tasks. To address privacy and communication constraints, federated learning (FL) has been adopted for distributed multi-modal model training, where only model parameters instead of raw data are uploaded. Traditional multi-modal FL methods face performance degradation due to the intricate interplay of three types of modality heterogeneity: modality quantity, modality feature emphasis, and modality statistical distribution. In this paper, we propose HeteroPMMFL, a novel personalized multi-modal FL framework. HeteroPMMFL proposes a model decoupling approach to effectively address modality combination differences and designs personalized collaboration graphs to strengthen cooperation among similar clients. This framework is applicable to any multi-modal dataset and can be easily extended to accommodate various modality combinations. Experimental results show that HeteroPMMFL outperforms both traditional FedAvg and the state-of-the-art Harmony framework in model accuracy, under the presence of three types of modality heterogeneity.

Cyber-physical systems (CPSs) are increasingly threatened by stealthy false data injection (SFDI) attacks, which compromise system integrity by manipulating control signals and introducing false sensor data. These attacks are particularly challenging due to their diversity and often indistinguishable nature. In response to this issue, our work uncovers the fundamental causes behind SFDI attacks in linear time-invariant (LTI) systems and elucidates the principles enabling their stealth. We present a novel virtual extended system framework designed to eliminate strictly stealthy attacks within the entire CPS. Utilizing deep reinforcement learning (DRL) methodologies, we pioneer the use of detection results for real-time SFDI attack classification. Through numerical simulations, we validate our proposed method’s effectiveness, demonstrating a classification accuracy of no less than 95%. Notably, even in scenarios where attackers manage to breach the framework partially, our method continues to provide a reliable success rate in SFDI attack detection and classification, showcasing its robustness and efficacy.Note to Practitioners—Cyber-physical systems (CPSs), a critical component of modern industries, are becoming increasingly susceptible to stealthy false data injection (SFDI) attacks. These attacks compromise system integrity by subtly manipulating control signals and feeding false sensor data, making them challenging to detect. Our research presents an innovative framework that uses deep reinforcement learning techniques to detect and classify these elusive attacks, achieving a classification accuracy of over 95%. The information on SFDI attack categories, ascertained by this method, lays the groundwork for the development of subsequent defence strategies. For professionals working in sectors reliant on CPSs, such as manufacturing, healthcare, and transportation, this framework offers a promising tool to enhance system security. Even in scenarios where the system has been partially compromised, our method continues to provide reliable detection and classification, underscoring its robustness and practical utility. The system remains effective despite full breach attempts on specific attack types, ensuring resilience against a broad range of SFDI attacks. In conclusion, our research offers a substantial advancement in protecting CPSs against cyber threats.

In the past decade, deep neural networks have been widely introduced into remote sensing image interpretation tasks. Exploring how to improve the model at some critical points, such as making it more suitable for remote sensing data, is still lacking, which makes more sense than developing stacked and transformed network structures, as is the expertise of scholars in the field of deep learning. To tackle this issue, this article proposes a fixed-length sequence-vision reformer (FLS-ViR) model. Relying on an advanced attention encoder with a reversible residual network structure, the presented model drastically decreases the spatial complexity. In contrast to most models that take input image sizes with 200+ pixels as edges, it will help us allow the input of large-size images under limited equipment conditions, thereby making fuller use of high spatial resolution data information. The introduction of a self-supervised method further improves the detection accuracy. Meanwhile, we built a high spatial resolution aerial image dataset for scene classification, which is superior in terms of high interclass similarity and diversity of image variations. To validate and evaluate the method and dataset, we performed a comparison of accuracy and memory consumption on a similar aerial dataset and attention-base models. The effectiveness of our model and its fitness on remote sensing datasets with a limited sample size is proved. Meanwhile, the investigation of the heat map demonstrates that increasing the size of the input image has a positive effect on optimizing the features captured by the model for remote sensing image scenes.

Convolutional neural networks (CNNs) have achieved success in HSI classification, but the performance is constrained by the limited reception field. In this regard, vision transformer is introduced recently, which is of powerful capabilities in long-range feature extraction for HSI classification. However, transformers are computation intensive and poor for local feature extraction. The motivation for this study is to build a lightweight hybrid model, which ensembles the respective inductive bias from CNNs and global receptive field from transformers. In this work, we propose a concise and efficient framework—the spatial-spectral feature extraction network with patch attention module (PASSNet), to simultaneously extract both local and global features. Specifically, we design an innovative plugin called patch attention module (PAM), which can be easily integrated into both CNNs and transformers blocks to extract spatial-spectral features from multiple spatial perspectives. Besides, a novel partial convolution operation is introduced, with a reduced computational cost than vanilla convolution operation. Through coupling the local attention from the CNNs with the global receptive fields in the transformers, the proposed PASSNet exhibits a superior classification performance on three well-known datasets with a small training sample size.

Recently, there is an emerging trend to apply deep reinforcement learning to solve the vehicle routing problem (VRP), where a learnt policy governs the selection of next node for visiting. However, existing methods could not handle well the pairing and precedence relationships in the pickup and delivery problem (PDP), which is a representative variant of VRP. To address this challenging issue, we leverage a novel neural network integrated with a heterogeneous attention mechanism to empower the policy in deep reinforcement learning to automatically select the nodes. In particular, the heterogeneous attention mechanism specifically prescribes attentions for each role of the nodes while taking into account the precedence constraint, i.e., the pickup node must precede the pairing delivery node. Further integrated with a masking scheme, the learnt policy is expected to find higher-quality solutions for solving PDP. Extensive experimental results show that our method outperforms the state-of-the-art heuristic and deep learning model, respectively, and generalizes well to different distributions and problem sizes.

We present a novel deep reinforcement learning method to learn construction heuristics for vehicle routing problems. In specific, we propose a Multi-Decoder Attention Model (MDAM) to train multiple diverse policies, which effectively increases the chance of finding good solutions compared with existing methods that train only one policy. A customized beam search strategy is designed to fully exploit the diversity of MDAM. In addition, we propose an Embedding Glimpse layer in MDAM based on the recursive nature of construction, which can improve the quality of each policy by providing more informative embeddings. Extensive experiments on six different routing problems show that our method significantly outperforms the state-of-the-art deep learning based models.

We present NeuroLKH, a novel algorithm that combines deep learning with the strong traditional heuristic Lin-Kernighan-Helsgaun (LKH) for solving Traveling Salesman Problem. Specifically, we train a Sparse Graph Network (SGN) with supervised learning for edge scores and unsupervised learning for node penalties, both of which are critical for improving the performance of LKH. Based on the output of SGN, NeuroLKH creates the edge candidate set and transforms edge distances to guide the searching process of LKH. Extensive experiments firmly demonstrate that, by training one model on a wide range of problem sizes, NeuroLKH significantly outperforms LKH and generalizes well to much larger sizes. Also, we show that NeuroLKH can be applied to other routing problems such as Capacitated Vehicle Routing Problem (CVRP), Pickup and Delivery Problem (PDP), and CVRP with Time Windows (CVRPTW).

Routing problems are very important in intelligent transportation systems. Recently, a number of deep learning-based methods are proposed to automatically learn construction heuristics for solving routing problems. However, these methods do not completely follow Bellman's Principle of Optimality since the visited nodes during construction are still included in the following subtasks, resulting in suboptimal policies. In this article, we propose a novel step-wise scheme which explicitly removes the visited nodes in each node selection step. We apply this scheme to two representative deep models for routing problems, pointer network and transformer attention model (TAM), and significantly improve the performance of the original models. To reduce computational complexity, we further propose the approximate step-wise TAM model by modifying one layer of attention. It enables training on larger instances compared to step-wise TAM, and outperforms state-of-the-art deep models with greedy decoding strategy.

Without maintaining the internal constraint of the received data, conventional two-step passive localization methods are considered to be suboptimal. Following the thought of direct processing, a novel localization method based on the information of time-difference-of-arrival and angle-of-arrival (AOA) is proposed in this letter. Similar to other direct localization algorithms, the proposed method does not require the procedure of data association. By taking advantages of the spatio-temporal processing, the proposed method can handle the underdetermined scenario (i.e., the number of emitters exceeds the number of sensors from all the stations) without the prior knowledge about the number of emitters. Compared with the localization algorithms based on AOA only, the proposed method has superior performance on the condition of low signal-to-noise ratio and large bandwidth.