With the widespread adoption of intelligent devices such as Augmented Reality (AR) glasses and Unmanned Aerial Vehicles (UAVs) in power equipment inspection, data privacy and efficient on-device inference becomes increasingly important. Federated Learning (FL) offers a solution by enabling collaborative model training without sharing raw data. However, naive model aggregation in FL often suffers from inefficiencies due to data and device heterogeneity, and the resulting models are typically too large-scale for deployment on resource-constrained terminals. To address these challenges, we propose FedCSAD, a hierarchical FL framework that combines Contextual Bandit-based Client pair Selection (CBCS) and Confidence-weighted Multi-teacher Knowledge Distillation (CMKD). CBCS optimizes aggregation by adaptively selecting high-quality model pairs using contextual information, while CMKD compresses global and edge models into compact student models suitable for terminal deployment. Experiments on CIFAR-10 and InsPLAD, a dataset based on real industrial data, demonstrate that FedCSAD outperforms existing methods in model accuracy, convergence speed, and deployment efficiency, offering a scalable and privacy-preserving solution for real-time intelligent inspection in the power industry.
As quantum computing poses a significant threat to traditional cryptographic systems, the need for post-quantum cryptographic (PQC) solutions has become urgent. Implementing these lattice-based algorithms on resource-constrained platforms, such as embedded systems and IoT devices, presents challenges. These challenges stem from the high computational and memory demands of the algorithms. Existing hardware designs often focus on accelerating specific computational tasks, like the Number Theoretic Transform (NTT), but are usually tailored to individual algorithms, making them inefficient for multi-algorithm systems. In this work, we address these challenges by designing KiRa: A Unified Memory Architecture for PQC that supports both CRYSTALS-Dilithium and CRYSTALS-Kyber at all NIST security levels. The architecture allows core components, such as Keccak, NTT, and sampling modules, to be shared across both Dilithium and Kyber algorithms. A uniform control logic and dataflow enable seamless switching between the two algorithms without duplicating resources. To support high-throughput polynomial operations, KiRa includes a dual-port memory design that efficiently handles polynomial data, and features a high-throughput NTT engine compatible with both algorithms. We implement and test the design on a Zynq UltraScale+ MPSoC FPGA platform. The results demonstrate a significant reduction in logic usage, memory access frequency, and dynamic power consumption, while improving latency for signing and decryption operations. Our results show that KiRa achieves a 23.8
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 .
With an increasing volume of data being collected and processed on edge, Federated Learning (FL) has become a dominant approach to train shared global models on distributed data held by diverse edge clients without exposing privacy. However, with expectations of high-quality user experience, instantaneous intelligent decisions must be made at the network edge to meet performance, security, and latency requirements. Intuitively, this challenge becomes more daunting as edge devices are constrained by limited computing and storage resources. Furthermore, edge clients need to continuously acquire, update, accumulate, and apply knowledge in domains requiring the processing of sensitive data and long-term model training. It has been proven that edge clients often forget previous tasks knowledge after learning new tasks, resulting in catastrophic forgetting. To mitigate this, we propose a Federated Meta Continual Learning (FMCL) approach. It requires edge clients not only to learn new tasks quickly but also to memorize rapidly. Specifically, we adopted a MAML (Model-Agnostic Meta-Learning) framework with memory replay, setting up buffers at edge clients to consolidate old tasks. In such a way, the model learns optimal initialization to quickly adapt to all seen tasks while preventing catastrophic forgetting. Extensive evaluation results indicate that FMCL performs well in alleviating catastrophic forgetting and achieves a good balance between learning new tasks and remembering existing tasks.
With the increasing concern about network security and model interpretability, fake audio detection is attracting significant attention. Although prior studies primarily concentrate on distinguishing authentic from fabricated audio, a vital yet underexplored dimension is the pinpointing within such detections, particularly in identifying the impersonated individual from counterfeit audio. In this paper, we propose a model capable of simultaneously detecting fake audio and identifying the speaker by extracting ID-specific features. Furthermore, we produce the Fake In-the-Wild Audio (FIA) dataset by expanding the "Inthe-Wild" audio dataset. We adopt the advanced Text-to-Speech generation model, MetaVoice-1B, to generate fake audios based on the "In-theWild" audio dataset. We conducted a detailed analysis of MetaVoice1B, focusing on its capabilities in generating realistic deepfake audio. Additionally, we identified its disadvantages and suggest potential future improvements. Experiments on the FIA dataset demonstrate the excellent performance of the proposed model, achieving a remarkable F1 score of 0.99 on the test set. Moreover, the model's ability to output additional relevant information enhances overall cybersecurity by providing deeper insights into potentially fake content.
Vehicular networks based on distributed machine learning have gained extensive research attention due to their outstanding performance compared to traditional methods. Specifically, machine learning systems can output vehicle control signals based on inputs from the object detection models. However, it has been demonstrated that machine learning is susceptible to security issues, where adversarial samples may mislead models into producing incorrect outputs. In this paper, we propose a novel adversarial patch attack against transformer-based object detection models. Our attack can make the object of a specific class invisible to object detection models. To enhance the attack performance and transferability of the adversarial patches, we designed the Gradient Self-Ensemble module and Transformer Masking Matrix module. These modules obscure parts of the inputs to each encoder and decoder and aggregate the outputs of all decoders within the model to generate adversarial samples, significantly improving the success rates of black-box attacks. We successfully suppressed the model's inference capabilities and deceived multiple transformer-based object detection models. Our work highlights the vulnerability of object detection models in vehicular networks to adversarial patch attacks in both digital and physical domains.
Graph database management systems (GDBMS) can efficiently store and deal with massive graph data that reflect complex relations among entities. However, most existing works can not adapt well to the query requirements of large-scale complex graph data. In this paper, a prototype CIGraph is presented based on a compressed index, which preserves a high compression ratio and query performance on graph data. Through sufficient experimental evaluations on LDBC-SNB benchmark, CIGraph with compressed index outperforms JanusGraph in aspects of both compression ratio and query performance. Compared with JanusGraph, CIGraph occupies less than a third of the space to store the same graph data, and the acceleration coefficient for graph algorithms is up to 3X on average in CIGraph.
In cross-domain Internet of Things (IoT) networks, the centralized intrusion detection system faces the challenge of privacy security and performance bottlenecks during centralized data processing. Distributed federated learning frameworks enable the training of models without sharing private data, thereby providing a highly scalable, privacy-friendly solution for IoT intrusion detection. However, federated learning also faces the challenge of label distribution skew caused by non-independent and identically distributed (non-IID) data, as well as client selection and aggregation efficiency. To address the above challenge, we propose an enhanced IoT intrusion detection method based on contrastive federated learning, i.e., ID-CFL. Firstly, the introduction of the federated learning framework alleviates privacy concerns as users do not need to share private data. Secondly, we design an intrusion detection method based on contrastive federated learning. It achieves abnormal detection by analyzing the inherent feature similarity among data, reducing the reliance on explicit label information, and demonstrating improved robustness in scenes with label distribution skew. Lastly, we propose an efficient federated aggregation algorithm based on node correlation degree. Extensive experiments on three well-known datasets demonstrate that the ID-CFL model exhibits outstanding performance in accuracy and communication efficiency.
Personalized Federated Learning (PFL) has gained significant attention for its superior capability in customizing model training. In response to the potential privacy risks associated with PFL, researchers apply differential privacy techniques to personalized federated learning, which protects privacy by clipping and adding noise to parameters. Existing methods utilize a uniform clipping threshold for all clients, which significantly reduces model accuracy. This reduction arises because these methods overlook the diversity in model parameters. To enhance the accuracy of personalized models under privacy constraints, we propose an Adaptive Dimensional Differential Privacy framework for Personalized Federated Learning (FedADDP). The framework utilizes Fisher information matrix to evaluate the sensitivity of parameters, effectively differentiating personalized parameters tailored to individual clients from global parameters consistent across all clients. Global parameters are handled through global consistency regularization and GlobalRobust Loss to ensure stability across different clients. Furthermore, we propose an adaptive dimensional differential privacy mechanism, which dynamically adjusts the clipping threshold for each dimension using historical gradient information, thus mitigating the accuracy loss of the personalized model. Experiments on the FEMNIST, SVHN, and CIFAR-10 datasets show that FedADDP improves accuracy by 1.67 https://github.com/yyguo-xdu/FedADDP .
Given the time-intensive nature of manual fundus examination and its significant dependence on expert knowledge, the task of computer-aided multi-classification of diabetic retinopathy (DR) severity is of crucial importance. However, the application of multi-classification for DR severity remains limited by a scarcity of high-quality training data, despite its potential. Moreover, challenges such as suboptimal data quality, ambiguous pathological characteristics, and minimal dissimilarity among categories are prevalent. Further research on image preprocessing methods is necessary. This study proposes a new framework for diabetic retinopathy classification, PMADR-Net (Progressive Multiscale Attention Network for Diabetic Retinopathy). By applying progressive multiscale training and attention mechanisms, this framework is suitable for few-shot medical image classification. It could automatically shift its focus towards learning samples of pathological features that are more challenging to discriminate. Moreover, in order to mitigate the class imbalance in the dataset, this research employs a balanced sampling approach and a patch reconstruction technique to produce multiscale features, thereby facilitating the model’s accurate differentiation of each category. Our proposed model achieved a classification accuracy of 93.97
As the Noisy Intermediate-Scale Quantum era progresses, the threat of cyberattacks using large-scale quantum computers to decrypt TLS communication becomes feasible. Fortunately, multiple contributions from the cybersecurity community to the National Institute of Standards and Technology's open call ensure the standardization of post-quantum algorithms that non-quantum devices can use to defend against such attacks. Various hardware and software implementations have been explored at each phase of the open call to identify potential threats and evaluate key performance metrics, such as CPU usage and RAM footprint. In this context, our research design and propose a Power Monitoring Framework that enables the monitoring of power constraints of an infrastructure that supports NIST's digital signatures and key establishment mechanisms. The proposed framework enables the benchmarking of power consumption and related metrics in both classical and Post-Quantum real-world infrastructure, contributing to the exploration of the Post-Quantum era's performance, requirements and constraints.
The development of VANETs brings hope for intelligent transportation. However, due to the dynamic and random feature of vehicles, the topology of the VANETs is constantly changing, making it vulnerable to hacker attacks that can lead to vehicle congestion, car accidents and even more serious consequences. The most notorious attacks are Black-Hole Attacks (BHA) and GrayHole Attacks (GHA). There are usually four variants of them: routing BHA, Intelligent BHA, Node-Dependent GHA, and Time-Dependent GHA. For these four types of attacks, this paper proposes a lightweight detection scheme that divides the detection process into three stages, using the proposed RR threshold algorithm, intersection algorithm, and forwarding rate-based algorithm to effectively detect malicious vehicles through the cooperation ofVehicles and Road Side Units (RSUs). This approach offloads computational tasks to the network's upper layers, thereby reducing the computational load on vehicles. Performance analysis results demonstrate that the algorithm is highly efficient and broadly applicable.
Due to the free and open source nature of the Android operating system, the number of Android malware is growing exponentially, which poses a serious threat to the property and privacy of Android users. Existing machine learning methods suffer from complex feature engineering, high workload, and weak generalization ability. In this paper, we propose WT-NET, a machine-learning based approach for Android malware detection, which first characterizes Android application traffic as a grayscale graph and transforms the traffic detection problem into an image classification problem. For the grayscale map characterization results, we further extract the time-frequency features of the traffic grayscale map using wavelet transform and achieve effective Android malware detection by combining the time-domain features with the frequency-domain features. To demonstrate the validity of WT-NET, we conducted an experimental evaluation using the publicly available dataset CICAndMal2017. Experimental results show that the method exhibits good performance in terms of efficiency and accuracy. Specifically, it was able to achieve 97.66
Deep Reinforcement Learning (DRL) has achieved significant progress in various domains, notably in autonomous systems and strategic gameplay, exemplified by developments such as AlphaGo and the MuZero algorithm series. Despite these advancements, the considerable computational demands of DRL models present challenges for their practical implementation in real-world applications. This research paper addresses the issue of hyperparameter prediction within the context of DRL training. We introduce HGAPerf, a method designed to forecast the run time during the training of DRL models. HGAPerf integrates heterogeneous graph hyperparameter mapping with predictive modeling, employing Graph Neural Networks (GNNs) enhanced by multi-head attention mechanisms. This approach aims to improve the accuracy of training time predictions for the MuZero series algorithm. To evaluate the performance of HGAPerf, we conducted a series of experiments, including baseline comparisons and an ablation study. The results indicate that HGAPerf can improve runtime prediction accuracy and enhance training stability. In comparative performance evaluations, HGAPerf showed improvements over existing models such as DNNPerf and Population Based Training (PBT). The model achieved a Mean Relative Error (MRE) of 13.4
Deep learning(DL) systems exhibit multiple behavioral characteristics such as correctness, robustness, and fairness. Ensuring that these behavioral characteristics function properly is crucial for maintaining the accuracy of DL systems' outputs. As a specialized form of software, DL systems' security testing techniques have increasingly become a focus of research in quality assurance. We analyze and organize the testing techniques for DL systems based on an investigation of the current state of the art both domestically and internationally. This paper categorizes existing approaches as component-oriented and attribute-oriented methods, providing a detailed review based on this classification. Additionally, we forecast the future development of testing techniques for DL systems.
With the advancement of the Internet of Things (IoT) technologies, there has been a rapid increase in the volume of IoT data, leading to escalating costs in storage, transmission, and analytics. The benefits of conventional data deduplication schemes are diminishing when applied to IoT data that is similar but distinct, necessitating the development of new approaches to accommodate these new scenarios. This paper proposes a deduplication and approximate analytics scheme for encrypted IoT data in fog-assisted cloud storage. The scheme is based on Generalized Deduplication (GD), Message Locked Encryption (MLE), Homomorphic Encryption (HE), and ciphertext conversion techniques. We employ GD to divide similar but distinct IoT data into bases and deviations, and perform deduplication on the encrypted base to achieve efficient storage while protecting data privacy. Additionally, we utilize Hybrid Homomorphic Encryption (HHE) techniques to convert the symmetric ciphertext of IoT data into homomorphic ciphertext, facilitating approximate analytics while ensuring privacy protection of IoT data in fog-assisted cloud storage and reducing the computation overhead on IoT devices.
Euphemism, which is often used to disguise true intentions or soften the tone of speech, has garnered significant attention for its detection in recent years. Current approaches to euphemism detection struggle with uncommon data, where new and rare vocabularies frequently emerge in the test set, leading to decreased model performance. To address these challenges, we propose a novel framework that combines data augmentation, self-training, and a special sample selection method. We expand the original dataset using lexical and sentence-level data augmentation techniques, enhancing the representation of rare vocabulary and allowing the model to familiarize itself with their contexts and usage characteristics. Next, we employ self-training as a semi-supervised learning approach, utilizing the augmented datasets to iteratively generate and refine pseudo-labels. During each iteration, a KNN framework is used to generate pseudo-labels by searching for the nearest neighbor texts, thus enhancing the quality of these labels. Finally, to further increase the presence of rare vocabulary in the training data, we implement a sample selection strategy based on euphemism frequency and predictive uncertainty, integrating it with the earlier steps. Experimental results demonstrate that our framework outperforms previous research on most metrics and achieves a 10
The complexity and unpredictability of disaster areas pose significant challenges to post-disaster rescue operations. UAVs, with their flexibility and rapid response capability, are able to perform various tasks and provide critical field data. However, these tasks are often computationally intensive and have strict latency requirements, and UAVs have limited computational resources. Therefore, effectively handling a large number of real-time tasks becomes a significant challenge. To address this, we propose a solution combining UAVs and mobile edge computing to optimize task processing. We focus on UAV-assisted task computation in post-disaster scenarios, considering indivisible and latency-sensitive tasks, and construct an edge computing framework for task offloading. We propose a deep reinforcement learning-based task offloading algorithm (TODRL), which predicts UAV loads using LSTM and adjusts offloading strategies with Dueling DQN to minimize latency. Experimental results show significant latency reduction compared to traditional methods, validating the approach's potential and effectiveness in postdisaster rescue.
In recent years, contrastive learning has become a prominent approach in unsupervised learning. Based on its powerful feature extraction capabilities, contrastive learning can make full use of the value of a large amount of unlabeled high-quality data for model training, thereby improving model quality. However, contrastive learning still faces serious privacy leakage issues when used in downstream tasks. Although recent research attempted to address this issue by incorporating differential privacy into model training, these approaches still suffer from low model utility, especially in the scenarios of extracting accurate feature by contrastive learning. In this paper, we propose DP-CLMI, a novel methodology by integrating differential privacy into the feature output layer of the contrastive learning model which acts as a feature extractor and is then used by untrustworthy parties for downstream tasks, thereby safeguarding the privacy of the training data. Our experimental results demonstrate that this method not only maintains a robust guarantee of differential privacy, but also preserves acceptable levels of classification accuracy. Additionally, the methodology proves to be resilient against various forms of membership inference attacks, especially neural network-based MIA.
Language-based colorization generates realistic and aesthetically appealing colors by leveraging the guidance of intuitive and userfriendly natural language descriptions. Previous methods in languagebased image colorization face several significant challenges, including limited color richness, color bleeding, color distortions, and inconsistent color styles. These issues often arise due to the reliance on dense attention mechanisms, which can lead to an overemphasis on global features at the expense of local details. Furthermore, many feature fusion modules for text and grayscale images are not designed to achieve adequate alignment of the diverse feature representations, resulting in suboptimal colorization outcomes. In this paper, we explore a novel module that employs sparse attention to mitigate the issue of color bleeding and limited colorfulness in image colorization tasks. By introducing a sparse attention mechanism, our method enables more flexible allocation of computational resources with a focus on color awareness. Additionally, we propose a module that effectively aligns grayscale images with color descriptions, thereby improving the consistency and quality of the colorization results. This module leverages the synergy between image features and language descriptions to solve long-standing issues like color bleeding. Empirical evaluations demonstrate that our approach surpasses recent state-of-theart techniques in both automatic and language-based colorization tasks, validating the effectiveness and robustness of our proposed method in generating high-quality, visually appealing colorized images.
