Deep learning-based website fingerprinting has emerged as an effective technique for inferring the websites users visit. Although existing methods achieve strong performance on closed-world datasets, they often fail to generalize to real-world environments, especially under geographic and temporal shifts. This limitation fundamentally stems from the coupled effects of two key challenges: application-layer resource composition variability and observable feature instability induced by cross-layer encapsulation. Intertwined, these factors induce systematic shifts between underlying application semantics and observable traffic features. To address the above challenges, we propose SATA , a semantics-aware traffic augmentation framework. Specifically, SATA first performs application-layer semantic augmentation based on protocol rules, expanding the resource composition patterns within each flow and frame sequence patterns under protocol constraints. Based on these augmented frame sequences, we further introduce a cross-layer feature alignment mechanism via knowledge distillation. It aligns frame sequence with packet-length sequence features, enabling cross-layer feature alignment between enhanced semantics and observable sequences. Extensive experiments show that SATA successfully generates traffic patterns that are absent from the training set but genuinely exist in the test set, and significantly improves the performance of mainstream models across diverse and complex scenarios. In particular, in open-world settings, SATA improves ACC by 90.81

Multimodal Named Entity Recognition (MNER) is an important research direction in Natural Language Processing, which aims to enhance Named Entity Recognition (NER) performance via additional image information. As posts with multiple images become increasingly common on social media, existing methods primarily focus on single-image scenarios, leaving a significant research gap for multi-image MNER. Current multi-image approaches often fuse holistic image representations with text, a strategy susceptible to noise from irrelevant visual information and which overlooks fine-grained, object-level cues crucial for entity recognition. To address this challenge, we propose a novel Object-level Semantic Alignment Framework. The framework utilizes an object detector to decompose the multiple images into a set of fine-grained candidate visual objects. A semantic alignment module then calculates the semantic relevance of each object to the text, based on which it selects the Top-K most critical visual proofs from the candidates, which are then fed into a multimodal interaction module for deep fusion with the text. Extensive experiments on public multi-image MNER datasets demonstrate that our proposed method significantly outperforms existing baselines. Ablation studies further validate the effectiveness of our object alignment and selection mechanism, providing a novel solution for the MNER-MI task.

Hierarchical federated learning (HFL) has emerged as a promising paradigm for distributed machine learning in edge computing environments. By introducing intermediate edge servers (ESs) between clients and the central server, HFL enhances system scalability and reduces communication latency. However, the heterogeneity of Internet of Things (IoT) devices-encompassing computational resources, data distributions, and network conditions-poses significant challenges for effective model training. Resource-constrained devices may fail to complete training tasks within designated time windows, resulting in wasted computational efforts, discarded learning rounds, and degraded model performance. To address these challenges, this article proposes HFed-CMS, a novel HFL framework that integrates intelligent client clustering, predictive selection, and fairness mechanisms. Our approach first performs hierarchical clustering (HC) based on the similarity of clients' local model weights, which implicitly captures underlying data distribution patterns and reduces intracluster heterogeneity. Subsequently, we introduce a predictive client selection strategy that evaluates multidimensional device capabilities-including central processing unit (CPU), memory, energy, and network bandwidth-to ensure that selected clients can successfully complete their assigned training tasks. Furthermore, we incorporate a fairness mechanism through bandwidth reutilization to guarantee equitable participation opportunities for all clients in long-term training scenarios. Through comprehensive experiments on multiple benchmark datasets under various heterogeneous settings (both system and statistical heterogeneity), we demonstrate that HFed-CMS significantly outperforms existing approaches.

Few-shot Multimodal Aspect-Level Sentiment Classification (MALSC) has garnered significant attention due to the proliferation of multimodal information on social media. In-Context Learning (ICL) with Large Vision-Language Models (LVLMs) offers a promising paradigm for MALSC. However, standard ICL is severely hampered by the high token cost of images, which restricts the number of contextual demonstrations. Circumventing this by replacing images with text captions often leads to critical information loss and error propagation. To address these limitations, we propose a novel framework based on Implicit In-Context Learning (I2CL). Our method compresses multiple image-text demonstrations into a single, compact “context vector” that encapsulates the essence of the task. During inference, this vector is directly injected into the LVLM’s activation space, guiding its reasoning without consuming any input tokens. This approach not only enables the model to leverage rich, raw visual information but also fundamentally resolves the token bottleneck and caption-induced errors inherent in previous methods. Extensive experiments on the Twitter-2015 and Twitter-2017 benchmarks demonstrate that our I2CL framework significantly enhances both performance and stability for few-shot MALSC, outperforming state-of-the-art models.

The performance of Federated Learning (FL) degrades significantly under Non-IID data distributions, severely constraining its applicability. Although Personalized Federated Learning (PFL) is a prevalent solution, it faces critical robustness challenges in practical, untrusted environments. Specifically, in decentralized settings, PFL is vulnerable to poisoning attacks via crafted gradients from malicious clients, which compromise gradient clustering precision and overall model accuracy. To address this, we propose a dual-layer dynamic reputation mechanism. By integrating real-time client supervision with decentralized aggregation verification, our framework guarantees fairness in cluster selection and model aggregation. Furthermore, it effectively mitigates the dominance of high-reputation nodes and alleviates the risks associated with centralization. Experiments demonstrate that the proposed method can effectively resist various types of attacks; specifically, under gradient tampering attacks, the model accuracy is improved by approximately 15% compared with the Fair method and achieves 98.5% accuracy on MNIST and 87.7% on CIFAR-10, outperforming methods including FedHKD, FedAGA, and BlockDFL.

The emergence of federated learning as a novel distributed machine learning paradigm has found widespread applications in the field of data security. However, the model aggregation mechanism inherent in federated learning relies on centralised devices, and while the integration of blockchain technology effectively mitigates this issue, it simultaneously introduces several challenges. In practical scenarios, federated client devices have different performance levels and latencies, and the incorporation of blockchain's consensus process significantly hinders the training efficiency of federated learning. Furthermore, clients are often untrustworthy and may exhibit malicious behaviour. To address these challenges, this paper proposes a practical and efficient Blockchain-based Dynamic Multi-Tier Federated Learning (B-DMTFL) framework. We consider the heterogeneity of clients in real-world network scenarios, and design a Dynamic FL computation scheme that facilitates reliable and efficient training within the federated learning process. At the same time, we integrate consistent hashing algorithm to formulate a new consensus protocol, thereby improving both the efficiency of consensus and the robustness of model aggregation. Extensive experiments demonstrate that this framework achieves accelerated convergence in applications that integrate blockchain with federated learning, while leveraging the blockchain's verification mechanism to ensure secure model aggregation and maintain high model accuracy.

The Android ecosystem faces persistent and rapidly evolving malware threats. Existing machine learning detectors are vulnerable to concept drift because they rely on implementation-specific features whose distributions change over time. Large language models (LLMs) offer strong semantic understanding and zero-shot reasoning, but current LLM-based detectors typically depend on code-centric or single-dimensional evidence, making them susceptible to obfuscation and limiting comprehensive behavior analysis. We present , a multimodal agent collaborative framework for dynamic Android malware detection. dynamically collects multimodal runtime evidence and employs ReAct-based specialized agents to analyze complementary behavioral views. The detection process begins by identifying visual deception cues, modeling UI state transitions, and integrating runtime API behaviors to fuse multi-dimensional evidence across user-visible interfaces and hidden backend operations. Experiments on temporally and distributionally unseen datasets show that achieves an accuracy of 90.06% without fine-tuning, outperforming state-of-the-art baselines and demonstrating strong zero-shot generalization against Android malware concept drift.

Federated Learning (FL) faces critical adoption barriers due to inherent centralized trust dependency and the lack of effective participant incentives. To resolve this dual challenge, we propose an innovative framework that integrates Blockchain technology with Game Theory. Blockchain establishes a decentralized and transparent infrastructure. Built upon this foundation, we utilize a Stackelberg Game model to create a dynamic incentive mechanism. The model’s utility function evaluates user contribution based on data quality and a comprehensive computation cost, including model training and Blockchain consensus overhead. This strategic interaction is formulated as a two-stage optimization problem, efficiently solved via Convex Optimization and Backward Induction. Simulation results confirm that our integrated scheme significantly enhances participant utility and resource allocation efficiency, providing a robust foundation for economically viable FL ecosystems.

Deduplication and data sharing are two crucial technologies designed to enhance the storage and sharing efficiency of cloud service providers. However, existing encryption-based solutions face critical challenges that hinder their effectiveness. (1) Some existing encryption algorithms introduce randomness, which reduces deduplication efficiency and hampers data sharing. (2) Other algorithms attempt to balance deduplication and data sharing but rely on third-party authority servers. To address these challenges, we introduce the SDS-Cloud scheme, which reduces storage overhead for cloud services and promotes the sharing of data. This scheme operates independently of third-party authority servers for both deduplication and data sharing while defending against various types of attacks. Specifically, deduplication is achieved using an optimized message-locked encryption scheme. Unlike traditional message-locked encryption, which introduces randomness, our approach maintains deduplication efficiency while preventing brute-force dictionary attacks. Proxy re-encryption is employed to manage data access rights, enabling data sharing without granting the cloud service provider access to the actual data. Moreover, security analysis shows that the proposed scheme meets the required security standards. Furthermore, simulation experiments demonstrate that our scheme outperforms existing approaches in both repetitive data verification and permission management efficiency.

Security vulnerabilities in smart contracts are a primary risk for Decentralized Finance (DeFi) and blockchain systems. Current research focuses on detection, neglecting end-to-end automated management. To address this, we propose LLM-BSCVM, an LLM-based framework for automated vulnerability management. We introduce a “Decompose–Retrieve–Generate” methodology that breaks down the workflow into six sub-tasks for specialized agents: detection, repair suggestion, risk assessment, repair, patch verification, and report generation. These agents collaborate in an automated chain, retrieving information from knowledge bases to enhance reasoning. Our evaluation shows LLM-BSCVM achieves an accuracy and F1-score over 91 https://github.com/sosol717/LLM-BSCVM .

Blockchain technology has driven the development of Decentralized Applications (DApps) in areas such as decentralized finance. However, as application scenarios become more complex, the limitations of computational resources and costs gradually lead to insufficient performance. Large Language Models (LLMs), as a promising technology, have the potential to enhance blockchain's capabilities in complex task governance. However, due to factors such as consensus mechanisms, it is challenging to directly integrate them with blockchain. To address this issue, this article proposes and implements a general framework for integrating LLMs with blockchain data, C-LLM, which successfully overcomes interoperability barriers between the two. By combining semantic relevance evaluation and truth discovery techniques, this article presents an innovative data aggregation method, SenteTruth, which effectively improves the correctness and credibility of data generated by LLMs. To validate the framework's effectiveness, we construct a dataset containing three types of questions, covering Q&A records between 10 oracle nodes and 5 LLM models. Experimental results show that, in the presence of 40% malicious nodes, the proposed method improves data correctness by an average of 17.74% compared with the optimal baseline. This research not only provides an innovative solution for the intelligent application of smart contracts but also demonstrates the potential for deep integration of LLMs and blockchain, driving the development of smarter and more complex application scenarios for smart contracts.

In federated fine-tuning of large language models (LLMs), commonly used aggregation methods include synchronous and asynchronous strategies. However, under the dual constraints of high-dimensional parameter spaces in LLMs and limited client resources, fully synchronous aggregation often leads to long waiting bottlenecks, while traditional asynchronous aggregation may suffer from staleness issues due to inconsistent client training progress, thereby affecting the convergence speed of the global model. To address this problem, we propose FedSLA, an efficient asynchronous federated fine-tuning method for LLMs based on time windows and staleness-controlled layer-wise aggregation. FedSLA accelerates the convergence of LLM fine-tuning in federated learning through a time-window-controlled asynchronous aggregation mechanism combined with explicit staleness control. In addition, it reduces communication overhead between clients and the server through a selective layer-wise aggregation strategy. We validate the feasibility and effectiveness of FedSLA in the context of LoRA-based fine-tuning for LLMs. Experimental results show that FedSLA’s layer-wise aggregation reduces communication time by 24% compared to full model aggregation; compared with fully synchronous aggregation, FedSLA shortens training time by 30.1%; and relative to traditional asynchronous aggregation, the staleness control mechanism improves global model accuracy by 4%, achieving convergence performance close to that of synchronous aggregation.

Reversible data hiding in encrypted images (RDHEI) is an essential data security technique. Most RDHEI methods with secret sharing cannot perform well on the images with low redundancy, such as the complex texture images. To address this issue, we propose an RDHEI method with ($n$, $k$) threshold-secret sharing (SS), which is universal for the images with diverse content since it is unrelated with the image content. Specifically, the original image is shared by polynomials over Galois field GF($2^{8}$) to generate $n$ image shares. Two kinds of reference matrices are elaborated to guide data embedding and extraction in image shares, one for higher capacity and the other for less communication overhead. At the decoder stage, the marked pixel shares are viewed as the unknowns to construct the multivariate linear equation (MLE) and the original image can be recovered by solving MLE. Experiment results show that the proposed method outperforms some state-of-the-art SS-based RDHEI methods.

Whole-system taint tracking is vital for security analysis. However, existing methods suffer from limited architecture compatibility and significant performance overhead, mainly due to the tight coupling between the functional and tracking domains. This paper introduces TaintEMU, an architecture-agnostic and efficient solution by fully decoupling the two domains. It separates functional and tracking logic at the QEMU TCG layer, mapping shadow registers to host instead of guest registers, ensuring compatibility across guest CPU architectures. At the host layer, it physically isolates the two domains: general-purpose instructions and registers serve the functional domain, while vector resources are dedicated to tracking, avoiding host resource reuse and enhancing tracking performance. Furthermore, it directly generates tracking instructions from TCG operations on the host, bypassing additional translation and further reducing overhead. We implement TaintEMU on an AMD64 host on QEMU 8.2.2. It supports a wide range of guest architectures (x86, MIPS, ARM, AMD, RISC-V, PPC), reduces performance overhead from 301% (DECAF++) to 101% and successfully detects all vulnerabilities in tests with 8 CVEs across 7 applications.

The rapid growth of large language models has increased the need for high-quality data, making efficient data acquisition a challenge. Blockchain-based crowdsourcing enhances data annotation efficiency in untrusted environments, but existing static or heuristic matching methods struggle in high-dimensional dynamic settings. This paper proposes an intelligent crowdsourcing framework using deep reinforcement learning for blockchain scenarios to improve data annotation in complex dynamic environments. The framework employs deep reinforcement learning for intelligent task-worker matching and dynamic allocation, optimizing scheduling based on task difficulty, worker performance, and resources. Blockchain smart contracts ensure secure, traceable, and fair user management, task management, and auditing. Experiments show the framework outperforms static and heuristic methods, improving on-time completion rate and transaction volume by 7.7

Mobile Edge Computing (MEC) enables Metaverse Terminal Devices (MTD) to perform complex tasks, including graphic rendering and physical simulation, by leveraging low-latency outsourced computing. However, existing research has not fully addressed the challenge of establishing an efficient outsourced computing service within an open and dynamic MEC environment that simultaneously ensures privacy and accountability. To address this, we proposes a blockchain-based privacy-preserving and accountable mobile edge outsourcing computing framework for the Metaverse, termed Meta-BMEOC. Specifically, we have designed an outsourcing computing protocol based on smart contracts and threshold secret sharing, enabling MTD to outsource tasks to multiple edge servers while preserving privacy. Furthermore, we have developed an off-chain smart contract protocol based on a Trusted Execution Environment. This protocol is designed to reduce the risk of malicious edge servers colluding to reconstruct the computational tasks of MTD, and it enables accountability for servers that return erroneous results. Additionally, we designed an incentive mechanism to resist malicious attacks and ensure system security and stability. Security analysis and experimental evaluation show that Meta-BMEOC not only ensures the privacy and accountability of outsourced computing but also provides outsourced computing services with lower computational latency.

Federated Learning (FL) allows participants to collaboratively train a global model without sharing local data. However, the non-independent and identically distributed (non-LID) nature of data in real-world scenarios leads to a decline in the accuracy of existing federated learning methods. Using Generative Adversarial Networks (LANs) to generate realistic pseudo-data is a feasible approach to alleviate the non-IID issue. However, the realism of GAN-generated data comes at the cost of diversity and is prone to falling unto local Therefore, this paper itiruwatively integrates the pretrained stable diffusion model's text -to-image mechanism with federated learning, constructing a client -server collaborative text generation mechanism that effectively alleviates the issue of data distribution bias. Furthermore, the effectiveness of FedDGIFID is validated on datasets such as CIFAR10 and CIFAR100.

Serious copyright infringement has become a critical issue in the widespread application of e-education and e-learning, as every link in the data flow has the potential to access complete data. In practice, entities that come into contact with the data may maliciously sell it, leading to copyright infringement. Previous efforts to prevent data leakage in intermediate circulation links have focused on access control, overlooking the copyright infringement of electronic education data by data requesters. To address these challenges, this paper proposes a traceability and accountability architecture for resource sharing, namely a malicious environment blockchain traceability architecture (MEB) based on multiple watermarks and oblivious transmission. The architecture first uses multiple watermarks and oblivious transmission to prevent false accusations and data leaks, then carries out evidence storage through the blockchain, third, finally carries out trials through the evidence on the blockchain after data leaks. We theoretically analyze the security of the scheme and verify its performance.

Encrypted traffic classification (ETC) is essential for ensuring network security and efficient management. Despite advances in deep learning, ETC remains challenging as existing models struggle to learn robust, discriminative representations from content-encrypted, highly imbalanced traffic.To address these challenges, we propose BTRFormer, a novel ETC approach that capitalizes on the inherent properties of encryption algorithms to enhance classification accuracy. At the core of BTRFormer lies a block-based, multi-layer traffic representation that adopts a 4×4 block as the fundamental unit, inspired by the encryption algorithm’s use of 16-byte blocks for encryption operations. This representation preserves the intrinsic structure of encrypted payloads, facilitating the model’s ability to learn deep semantic features. Subsequently, a transformer-based model is employed to learn from the multi-layer representation, capturing intra-block, inter-block, and inter-packet dependencies through block-wise attention mechanisms. Finally, BTRFormer leverages a pre-training phase on large-scale unlabeled data, followed by fine-tuning with a minimal amount of labeled samples to improve generalization and adaptability. Experimental results show that BTRFormer significantly outperforms SOTA methods on six real-world datasets, highlighting its effectiveness in encrypted traffic classification and secure network management.

Blockchain is distinguished by its decentralization and security, but it faces significant scalability challenges. Sharding is widely regarded as a key solution for improving blockchain scalability. However, existing sharding schemes, such as Monoxide, often suffer from load imbalance and excessive cross-shard transactions (TXs), which degrades system performance. To address these issues, we propose Sharding-aware NSGA-H (SNSGA-H), a variant of the Non-dominated Sorting Genetic Algorithm H, designed to optimize account allocation in blockchain sharding. We formulate the allocation problem as a multi-objective optimization task and develop a SNSGA-H-based sharding algorithm that dynamically adjusts account allocation to improve load balancing while minimizing cross-shard TXs. The proposed approach achieves a well-balanced trade-off in terms of Pareto efficiency, ensuring that improvements in load balancing do not come at the cost of excessive cross-shard TXs. We evaluate our method through simulation experiments on a real Ethereum dataseL The results demonstrate that our method significantly outperforms existing solutions, including Metis, Monoxide, and HyperChain, in key metrics such as throughput and TX confirmation delay.