Federated Learning (FL) is inherently vulnerable to model poisoning attacks owing to its distributed architecture. Existing defense mechanisms typically adopt a horizontal strategy—aggregating and processing all user gradients (model updates) within each communication round to derive optimal aggregated gradients. However, such horizontal approaches fail completely under large-scale poisoning attacks involving more than 50% of participants. In this work, motivated by the observation that model convergence follows a highly predictable trajectory, we depart from conventional horizontal paradigms and reformulate the problem of optimal gradient aggregation along a vertical dimension. We introduce VERT, a novel defense framework that leverages historical gradients from previous global rounds to train a predictor, which forecasts the expected gradients for the current round. These predictions are then compared against actual submitted gradients to identify and select reliable updates for aggregation. To enhance computational efficiency, VERT incorporates a lightweight dimensionality reduction module that projects high-dimensional gradients into a lower-dimensional space without compromising representational capacity. Extensive experimental results demonstrate that VERT is both efficient and scalable, outperforming state-of-the-art (SOTA) defenses even under extreme poisoning scenarios (≥80% malicious users). Furthermore, VERT consistently achieves superior robustness across low poisoning rates (20%, 40%) and non independent and identically distributed (non-IID) data settings.

One of the main challenges in federated learning is its non-independent and identically distributed (non-IID) nature, where independent client training leads to overfitting and model deviations, negatively impacting overall performance. To address this, most research focuses on aligning local and global models to reduce client drift. However, existing algorithms use Empirical Risk Minimization (ERM) as the local optimizer, leading the global model to sharp minima, increasing bias in some clients and lowering generalization. To address these challenges, we propose GFed-HSAM, a general federated learning method that improves both local and global model generalization. GFed-HSAM uses Hybrid Sharpness-Aware Minimization (HSAM) as the local optimizer to smooth gradients with zeroth-order and first-order sharpness measures. It also includes a dynamic regularizer (DR) to align global and local models at the parameter level. Experiments show that GFed-HSAM outperforms state-of-theart methods in accuracy and generalization across different data heterogeneity settings on CIFAR10/CIFAR100.

Federated Learning (FL) is a distributed training paradigm that enables multiple devices in the Industrial Internet of Things (IIoT) to collaboratively train a global model without sharing private data. However, non-IID data in FL leads to client drift, which significantly degrades the performance of the global model in IIoT scenarios. While the primal-dual update method effectively mitigates client drift through dynamic regularization, optimizing the global model remains a significant challenge in IIoT due to the high degree of data heterogeneity. To address this challenge, we propose a novel FL method, PDFed-ALD, which effectively mitigates client drift and improves global model’s performance under high data heterogeneity. The core of PDFed-ALD is adaptive local distillation mechanism, which employs an adaptive distillation temperature based on the relative degree of data heterogeneity, dynamically correcting gradient updates, alleviating the issue of client drift. Furthermore, to reduce variance among local gradients, PDFed-ALD introduces a momentum-based minimum sharpness gradient correction method, which enhances local consistency by minimizing the variance between gradients across clients. Extensive experiments on image classification tasks using CIFAR-10, CIFAR-100 and MVTEC datasets demonstrate that PDFed-ALD outperforms state-of-the-art (SOTA) methods in terms of both accuracy and convergence speed across various settings, including client scale, participation rate, and degree of data heterogeneity.

Large-scale storage systems boost I/O performance via flash caching, but the underlying storage medium of flash caching incurs significant financial costs and also exhibits low endurance. Previous studies adopt compression-after-deduplication to mitigate writing redundant contents into the flash cache, so as to address the cost and endurance issues. However, deduplication and compression have conflicting preferable cases, and compression-after-deduplication essentially compromises the space-saving benefits of either deduplication or compression. To simultaneously preserve the benefits of both approaches, we explore compression-before-deduplication, which applies compression to eliminate byte-level redundancies across data blocks, followed by deduplication to write only a single copy of duplicate compressed blocks into the flash cache. We present CDCache, a space-efficient flash caching system that realizes compression-before-deduplication. It proposes to dynamically adjust the compression range of data blocks, so as to preserve the effectiveness of deduplication on the compressed blocks. Also, it builds on various design techniques to approximately estimate duplicate data blocks and efficiently manage compressed blocks. Trace-driven experiments show that CDCache improves the read hit ratio and the write reduction ratio of a previous compression-after-deduplication approach by up to 1.3× and 1.6×, respectively, while it only has small memory overhead for index management.

Summary This article addresses the issue of ensuring model accuracy and training efficiency in a constrained federated learning environment. In an actual federated learning environment, each device's software, hardware, and network conditions are heterogeneous. Some terminal devices may not be able to undertake the work assigned by the server, resulting in poor model accuracy and slower convergence speed. However, existing research cannot ensure that each terminal device participating in training can handle the workload allocated by the system without collecting too much equipment information. This article proposes the cloud‐edge‐terminal collaborative self‐adaptive federated learning framework (CCSFLF) to solve this problem. This framework combines the advantages of federated learning and edge computing, reduces the probability that devices cannot handle the workload of system allocation, solves the system heterogeneity, and improves the efficiency of federated learning. CCSFLF can adaptively adjust the number of training tasks for terminal devices and select valuable training participants using a terminal device selection strategy. Multiple edge servers can simultaneously aggregate local models. Cloud servers are responsible for the aggregation and task distribution of global models. The above strategy enables the framework to have a faster convergence rate and higher model accuracy. The experimental results confirm that this framework can reduce the dropout rate of terminal devices by more than 5% in heterogeneous federated learning systems, improve the model accuracy by about 2%, and reduce the training time by 1/3 compared with similar methods, with better performance and applicability.

As a distributed learning paradigm , federated learning can be effectively applied to the decentralized system since it can resolve the “data island” problem. However, it is also vulnerable to serious privacy breaches . Although existing secure aggregation technique can address privacy concerns, they also incur significant additional computation and communication costs. To address these challenges, this paper offers a C ommunication E fficient S ecure A ggregation scheme. Firstly, the central server uses the communication delay between terminals as the weight of the fully terminal-connected graph to transform it into a sparse connected graph based on the minimal spanning tree. Secondly, instead of relying on central server for key advertisement , the terminals advertise keys via a neighboring terminal forwarding approach based on sparsely graph. Thirdly, we propose using the central server for auxiliary advertising to address unexpected terminal dropout. Simultaneously, we theoretically demonstrate our scheme’s security and have lower computation and communication costs. Experiments show that CESA can reduce the running time by 28.2% without sacrificing security and model accuracy compared to conventional secure aggregation when there are 10 terminals in the system.

We present Acoustic Inertial Measurement (AIM), a one-of-a-kind technique for indoor drone localization and tracking. Indoor drone localization and tracking are arguably a crucial, yet unsolved challenge: in GPS-denied environments, existing approaches enjoy limited applicability, especially in Non-Line of Sight (NLoS), require extensive environment instrumentation, or demand considerable hardware/software changes on drones. In contrast, AIM exploits the acoustic characteristics of the drones to estimate their location and derive their motion, even in NLoS settings. We tame location estimation errors using a dedicated Kalman filter and the Interquartile Range rule (IQR) and demonstrate that AIM can support indoor spaces with arbitrary ranges and layouts. We implement AIM using an off-the-shelf microphone array and evaluate its performance with a commercial drone under varied settings. Results indicate that the mean localization error of AIM is 46% lower than that of commercial UWB-based systems in a complex 10m×10m indoor scenario, where state-of-the-art infrared systems would not even work because of NLoS situations. When distributed microphone arrays are deployed, the mean error can be reduced to less than 0.5m in a 20m range, and even support spaces with arbitrary ranges and layouts.

Nowadays, microservice architecture has become mainstream way of cloud applications delivery. Distributed tracing is crucial to preserve the observability of microservice systems. However, existing trace representation approaches only concentrate on operations, relationships and metrics related to service invocations. They ignore service events that denotes meaningful, singular point in time during the service's duration. In this paper, we propose iTCRL, a novel trace contrastive representation learning approach based on causal intervention. This approach first constructs a unified graph representation for each trace to describe the runtime status of service events in traces and the complex relationships between them. Then, Causal-intervention-based Trace Contrastive Learning is proposed, which learns trace representations from causal perspective based on the unified graph representations of traces. It uses causal intervention to generate contrastive views, heterogeneous graph neural network-based trace encoder to learn trace representations, and direct causal effect to guide the training of trace encoder. Experimental results on three datasets show that iTCRL outperforms all baselines in terms of trace classification, trace anomaly detection, trace sampling and noise robustness, and also validate the contribution of Causal-intervention-based Trace Contrastive Learning.

Federated learning is a distributed learning paradigm, which is usually combined with edge computing to meet the joint training of IoT devices. A significant challenge in federated learning lies in the statistical heterogeneity, characterized by non-independent and identically distributed (non-IID) local data across diverse parties. This heterogeneity can result in inconsistent optimization within individual local models. Although previous research has endeavored to tackle issues stemming from heterogeneous data, our findings indicate that these attempts have not yielded high-performance neural network models. To overcome this fundamental challenge, we introduce the framework called FedPKR in this paper, which facilitates efficient federated learning through knowledge review. The core principle of FedPKR involves leveraging the knowledge representation generated by the global and local model layers to conduct periodic layer-by-layer comparative learning in a reciprocal manner. This strategy rectifies local model training, leading to enhanced outcomes. Our experimental results and subsequent analysis substantiate that FedPKR effectively augments model accuracy in image classification tasks, meanwhile demonstrating resilience to statistical heterogeneity across all participating entities. Code is available athttps://github.com/jbwangnb/FedPKR.

Edge intelligent computing is widely used in the fields, such as the Internet of Medical Things (IoMT), which has advantages, including high data processing efficiency, strong real-time performance and low network delay. However, there are many problems including privacy disclosure, limited calculation force, as well as scheduling and coordination issues. Federated learning can greatly improves training efficiency. However, due to the sensitive nature of the healthcare data, the aforementioned approach of transferring the patient's data to the servers may create serious security and privacy issues. Therefore, this article proposes a Privacy Protection Scheme for Federated Learning under Edge Computing (PPFLEC). First of all, we propose a lightweight privacy protection protocol based on a shared secret and weight mask, which is based on a random mask scheme of secret sharing. It is more accurate and efficient than,homomorphic encryption. It can not only protect gradient privacy without losing model accuracy, but also resist equipment dropping and collusion attacks between devices. Second, we design an algorithm based on a digital signature and hash function, which achieves the integrity and consistency of the message, as well as resisting replay attacks. Finally, we propose a periodic average training strategy, compared with differential privacy to prove that our scheme is 40 % faster in efficiency than in deferential privacy. Meanwhile, compared with federated learning, we can achieve the same efficiency under the condition of ensuring safety. Therefore, our scheme can work well in unstable edge computing environments such as smart healthcare.

The Internet of Medical Things (IoMT) is an important application of the Internet of Things in health care. In IoMT, efficiency and user privacy are crucial for cloud storage and retrieval of healthcare data documents. Existing schemes, however, often suffer from inefficient retrieval and increased risk of privacy disclosure when dealing with massive data. We propose here a new Efficient Encrypted Parallel Ranking (EEPR) search system, block-based and privacy-preserved, for encrypted cloud healthcare data. We design a parallel binary search tree structure in block and propose a parallel retrieval algorithm adaptable to such a structure. A quantitative analysis through the information retention index shows that our scheme demonstrates better search performance. In addition, feature vectors generated from our scheme are difficult to be reversely analyzed due to unexplainability, enhancing privacy protection for patients and researchers. A formal security analysis shows that our EEPR scheme is resistable to known background attack, and yields a lower time complexity and significantly improves search efficiency as well as accuracy over existing schemes.

We present Acoustic Inertial Measurement (AIM), a one-of-a-kind technique for indoor drone localization and tracking. Indoor drone localization and tracking are arguably a crucial, yet unsolved challenge: in GPS-denied environments, existing approaches enjoy limited applicability, especially in Non-Line of Sight (NLoS), require extensive environment instrumentation, or demand considerable hardware/software changes on drones. In contrast, AIM exploits the acoustic characteristics of the drones to estimate their location and derive their motion, even in NLoS settings. We tame location estimation errors using a dedicated Kalman filter and the Interquartile Range rule (IQR). We implement AIM using an off-the-shelf microphone array and evaluate its performance with a commercial drone under varied settings. Results indicate that the mean localization error of AIM is 46% lower than commercial UWB-based systems in complex indoor scenarios, where state-of-the-art infrared systems would not even work because of NLoS settings. We further demonstrate that AIM can be extended to support indoor spaces with arbitrary ranges and layouts without loss of accuracy by deploying distributed microphone arrays.

As traditional computing and cloud computing integrate, the Internet of Things (IoT) has evolved into a layered and cloud-network-edge-end architecture. However, most searchable encryption models still use triples, in which hierarchical structures are neglected, and insecure intermediate nodes are exposed to external environment. Meanwhile, mainstream schemes adopting accurate retrieval are incompatible with IoT end users' features of differentiation. To address these issues, we innovatively design an efficient and credible search model with an accurate multikeyword fuzzy ciphertext retrieval scheme in the context of IoT. First, based on network coding and key sharing, data are grouped, encoded, and transmitted in parallel to the receiver node through middle-layer nodes, with high efficiency and reliability. Second, to realize fuzzy retrieval of IoT, edit distance is selected as the standard of difference between keywords, and then document index vector and query vector are created based on locality sensitive hashing (LSH) and Bloom Filter. Furthermore, to improve the traditional scheme, query keywords are split into multiple single-word forms, inner products between each trapdoor of single word and encryption index vector are calculated, respectively, for the sum of each inner product and thus top $\mathrm{top}$ - k $k$ sorting search. Ultimately, feasibility, safety, and efficiency of our improved scheme are verified by security analysis, while simulation results support that our scheme has better accuracy and efficiency.

With development of online social networks, many criminal suspects use social network to communicate with each other. In order to obtain valuable criminal clues, considerable research works have been done to analyze criminal suspects' social data. However, most of them did not pay much attention on privacy-preserving problems, which may leak some sensitive data in the analysis process. To solve this problem, we propose a novel analysis approach of criminal suspects by exploiting social data and crime data that are collected by social network and police information systems. We enable the social cloud server and public security cloud server to exchange social information of criminal suspects and user's public information in a privacy-preserving way. Specifically, we propose a privacy-preserving data retrieving method based on oblivious transfer to guarantee that only the authorized entities can perform queries on suspects' social data, while the social cloud server cannot infer anything during the query. Moreover, several building blocks, such as encrypted data comparing, secure classification and regression tree (CART) model are also proposed. Based on these building blocks, we designed a privacy-preserving criminal suspects sensing scheme. Finally, we demonstrate a performance evaluation which shows that our scheme can enhance analysis of criminal suspects without privacy leakage, while with low overhead.

To improve the accuracy and automation of malware Uniform Resource Locator (URL) recognition, a joint approach of Convolutional neural network (CNN) and Long-short term memory (LSTM) based on the Attention mechanism (JCLA) is proposed to identify and detect malicious URL. Firstly, the URL features including texture information, lexical information and host information are extracted and filtered, and pre-processed with encode. Then, the feature matrix more relevant to the output are chose according to the weight of the attention mechanism and input to the constructed parallel processing model called CNN_LSTM, combinating CNN and LSTM to get local features. Next, the extracted local features are merged to calculate the global features of the URLs to be detected. Finally, the URLs are classified by the SoftMax classifier using global features, the accuracy of the model in malicious URL recgonition is 98.26%. The experimental results show that the JCLA model proposed in this paper is better than the traditional deep learning model or CNN_LSTM combined model for detecting malicious URLs.

As the indoor location-based services are widely used in daily life, we all call for precise localization and low energy consumption application. Besides, getting location information in a short re-sponse time remains a problem in indoor navigation. In this pa-per, we present an accurate and efficient indoor navigation sys-tem called AENS, which only needs available sensors in off-the-shelf smartphones such as accelerometer and gyroscope, completely avoid using energy hungry WiFi module. Combined with simple local map information, AENS is able to guide user to find destinations. The System utilizes those sensors to do Dead-Reckoning dynamically and in real-time, then fitting the predict-ed trajectories to the map information. We evaluate the system in our university library. The result shows that AENS successfully guides the tester to the specified destination, and for every critical node the average localization error is within one-meter.

Leveraging a wireless multipath channel as the source of common randomness, many key generation methods have been proposed according to the information-theory security. However, existing schemes suffer a low generation rate and a low entropy, and mainly rely on nodes' mobility. To overcome this limitation, we present a key generation protocol with known artificial interference, named SmokeGrenade, a new physical-layer approach for secret key generation in a narrowband fading channel. Our scheme utilizes artificial interference to contribute to the change of measured values on channel states. Our theoretical analysis shows that the key generation rate increases with the increment of the interference power. Particularly, the achievable key rate of SmokeGrenade gains three times better than that of the traditional key generation schemes when the average interference power is normalized to 1. Simulation results also demonstrate that SmokeGrenade achieves a higher generation rate and entropy compared with some state-of-the-art approaches.