Personalized Federated Learning (PFL) tackles the challenges of FL on heterogeneous data and provides customized solutions to each client. However, like commonly employed FL settings, PFL is still vulnerable to attacks on privacy and model availability. Existing PFL frameworks focus on either privacy protection or attack defense instead of simultaneously implementing both functionalities. To address this challenge, we design and implement a novel Privacy-preserving and Robust Personalized Federated Learning framework, PR-PFL, which can simultaneously protect data privacy and defend against model availability attacks. Specifically, PR-PFL adopts Mean Regularized MultiTask Learning (MR-MTL) as the base training paradigm. Clients perform per-sample DP to protect data privacy, and the central server executes a robust aggregation algorithm to filter out potential attackers. After collective training, clients tune their models locally to eliminate malicious injections further. To the best of our knowledge, this is the first PFL framework that protects both clients' privacy and model availability. The method combining robust aggregation and local tuning we have designed can effectively defend against 5 kinds of attacks. We conduct an extensive empirical evaluation demonstrating that our framework is practical and achieves reasonable robustness under an honest majority setting (attackers <50%).
In the Multi-sensor multi-vehicle (MSMV) cooperative localization system, the data incest problem generates inter-estimate correlation, leading to over-convergence problems. For the purpose of solving the data incest problem, we propose a method named interval split covariance intersection filter (ISCIF). First, the consistency proof of ISCIF is provided. Then, ISCIF is applied to achieve vehicle cooperative localization, including both absolute and relative positioning stage. Moreover, in order to reduce the computational overhead and address pessimistic estimation problems caused by multiple relative position estimates in the MSMV cooperative localization system, the interval constraint propagation (ICP) method is employed for data preprocessing before relative positioning using ISCIF. Finally, two different simulation experiment scenarios are designed to validate localization accuracy. Simulation results demonstrate that our proposed method, compared to the traditional split covariance intersection filter (SCIF) method, reduces the root mean square error (RMSE) by 13% and 21% in two different experimental scenarios, respectively.
Due to the pressing demand for advanced 6G and beyond communications technology that can effectively handle large-scale requirements with utmost security and efficiency. With the growth and spread of companies and institutions that deal with a large amount of data, protecting data and ensuring privacy has become a basic requirement. Actual reality and recent research have shown a significant increase in phishing scams. This research seeks to develop the security level of Li-Fi technology by detecting and preventing phishing activities. It is adopted to inspect URL links transmitted over the network before they reach the end user. The proposed system was tested using two-layer algorithms to detect and prevent hacking and phishing. The performance of the proposed system was examined on Li-Fi technology via FSO under different weather conditions. The proposed system has demonstrated ambitious results in implementing tasks successfully and effectively.
PointNet was a breakthrough in neural networks for directly processing point cloud data, ensuring transformation and input invariance with a simple, efficient structure. However, it neglects the relationships between individual points’ features. This paper introduces an improved model, KMeans-PointNet, which combines PointNet with K-Means clustering. By clustering point cloud data and extracting global features from each cluster, the model enhances local feature utilization. Experiments on the ModelNet40 dataset show that KMeans-PointNet improves overall recognition accuracy by 3% and classification accuracy by 2% compared to the original PointNet.
Various Internet of Things (IoT) devices bring great convenience to users. However, the privacy protection and overall performance still need improvements due to their limited storage and computation abilities. Cloud computing can be used to solve or lessen these issues by providing server-level security, unlimited storage, and high computation abilities. It is time-consuming to conduct all the calculation in local IoT devices. It is also not wise to offload everything to Cloud, because the data could be massive that takes longer time to transfer than to compute. In this paper, we design a framework in which the local IoT devices are the cache and the remote Cloud servers are the memory. To further improve system efficiency and adaptability, we use some lightweight Artificial Intelligence (AI) methods to predict users’ behavior patterns, that can adapt to changes while maintaining optimal performance.
In autonomous driving, the optimization theory and algorithms for distributed intelligent systems are essential for enhancing vehicle decision-making capabilities, path planning, and environmental perception. Evolutionary algorithms, as a global optimization method inspired by biological evolution, is widely used in these scenarios. Unfortunately, the performance of evolutionary algorithms relies on the properties of the solution space, the optimization strategy, and parameter settings, and their efficiency also requires more rigorous evaluation. This study focuses on differential evolution algorithm, examining the properties of the solution space for optimization problems, the preferences of strategies and parameters in different optimization scenarios, and setting new standards for efficiency evaluation. Based on the experimental conclusions, modifications are made to our algorithm to further verify the correctness of our findings. All experimental results are deduced based on the CEC2017 benchmark problem suite from the IEEE Congress on Evolutionary Computation (CEC).
In this paper, we consider a cooperative relay network consisting of a source, multiple intermediates, and a destination in the face of an eavesdropper. We propose a multiple friendly jammers (FJs) aided relay selection (MFJRS) scheme where an intermediate that maximizes the main channel capacity is chosen as the best relay, and the remaining intermediates act as FJs which can emit artificial noises to confuse the eavesdropper. We derive closed-form expressions of both outage and intercept probabilities and analyze the security-reliability tradeoff (SRT) performance for the MFJRS scheme. Numerical results show that the MFJRS scheme performs better than a conventional relay selection (CRS) and one FJ aided relay selection (OFJRS) schemes in terms of their SRT performance, indicating the effectiveness of multiple FJs in enhancing transmission security and reliability. It is also found that optimizing the power allocation factor between the best relay and multiple FJs further enhances the SRT performance.
Earthquakes are frequent global occurrences, impacting the safety of billions of people worldwide. Rapid transmission of seismic data enables monitoring agencies to promptly issue earthquake warnings, while secure transmission helps governments and media manage public perception, preventing panic and the spread of misinformation, thereby maintaining social stability and public safety. Consequently, the fast and secure transmission of seismic monitoring data has become a critical issue in recent years. To meet the high demands for timeliness and confidentiality in seismic information, this paper proposes a high-performance data packet processing method implemented on general-purpose platform to achieve low-latency transmission of seismic data. Additionally, encryption at the link layer is employed to ensure data confidentiality and support secure data transmission. This approach aims to ensure the accurate and at a low cost of seismic information, ultimately saving more lives during the critical moments following an earthquake.
In unmanned aerial vehicles(UAV) swarm combat, efficient management of electromagnetic spectrum resources is crucial to ensure the operational performance of UAV swarm. However, in the presence of external radiations, the spectrum interference and competition between UAV onboard communication, radar, reconnaissance, and jamming services greatly affect the combat effectiveness of UAV swarm. Besides, because of the heterogeneity and coupling of different services, traditional spectrum allocation methods lack flexibility and occupy many computing resources in the central nodes. In this paper, by setting the threshold and the utility function for each service, we model the multi-service spectrum scheduling as a joint utility maximization problem. Then, we propose an intelligent and lightweight spectrum management method based on deep Q-network for multi-service cooperative spectrum scheduling. By dynamically and collaboratively allocating spectrum resources, conflicts within and between services are greatly reduced. Extensive simulations demonstrate that the proposed approach provides robust and efficient multi-service spectrum management.
This paper presents and studies the expected bit error rate (BER) of a prepare-and-measure quantum key distribution (QKD) protocol utilizing two wavelengths, ternary data, and a physical unclonable function (PUF). Models are developed as a function of channel and PUF noise, with and without eavesdropping, and then are subsequently validated through software simulations. At channel error rates below 30% and PUF error rates up to 10%, an eavesdropper presents a detectable error rate without Alice and Bob having to reveal a sample of their key.
In recent years, wearable Internet of Things (IoT) devices, such as brain-sensing headsets, have gained popularity for their ability to provide insights into users’ well-being. These devices record Electroencephalography (EEG) signals, which are traditionally used in medical settings to monitor brain activity. With the advent of affordable EEG headsets, this technology has become more accessible to the general public. However, these devices are not without security vulnerabilities, which hackers can exploit to intercept users’ EEG data. Recent research has demonstrated that EEG signals, when processed and analyzed using machine learning, can be used to infer user conditions or activities, such as muscle movements, emotions, or even limited keyboard input.In this paper, we explore the potential of deep learning to infer the digits and words typed on a keyboard based on a person’s recorded EEG signals, achieving high accuracy. If deep learning techniques can be employed to deduce user activity from EEG signals, they could potentially expose sensitive and private information, such as password entry. Our contributions include demonstrating an attack model by developing and comparing various deep learning models to infer the digits and words typed by a person wearing an EEG headset. We also detail the process of reading raw EEG signals from a consumer-grade EEG headset and implementing a series of noise filters and signal transformations to refine the signal samples. Finally, we develop and test both CNN and RNN models for EEG signal classification to accurately infer passwords.
Detecting forged and generated images has recently grown into an emerging research area. As forgery and generation technologies advance, they pose risks of personal privacy and public security. Existing algorithms are designed to detect either forged or generated facial images. Due to a lack of generalizability, their performance usually degrades when faced with a mixture of both types. To tackle this problem, this paper proposes a framework Res50_Attn_DSCE that enhances generalizability and extracts both local and global features, thereby improving the algorithm’s performance across different types of datasets. Additionally, depth-separable convolution reduces computational costs. Experimental results demonstrate that our proposed model achieves the highest accuracy with the shortest runtime. Compared to other traditional algorithms, these results validate the effectiveness of our model’s improvements.
This paper presents RAS-Rec, a novel and robust approach to sequence recommendation that leverages the power of data augmentation techniques. RAS-Rec proposes data augmentation as a means to enhance the representations of users and items in sequence recommendation. By generating synthetic training samples through various transformations applied to the original data, RAS-Rec increases the diversity and quantity of the training set. The augmented sequence data addresses the limitations of existing methods and improves the robustness and generalization capabilities of sequence recommendation models. Additionally, a contrastive learning scheme is designed to train and further enhance the representations in RAS-Rec. Extensive experiments validate the effectiveness of the proposed approach, demonstrating its superiority over existing methods in terms of recommendation performance. The results highlight the value of data augmentation and contrastive learning in sequence recommendation tasks. The findings of this study contribute to the advancement of robust sequence recommendation techniques.
Human Activity Recognition was a key part of medical and mobile applications in today's world, with important applications in old-age homes and monitoring daily activities. In this paper, we proposed an automated Human Activity Recognition model using a 1-Dimensional Convolutional Neural Network (CNN). To select the optimal parameters for the CNN model, we used the Grid Search Cross Validation method. On one hand, we used a manual feature extraction and CNN method. Starting from the physical significance of accelerometer and gyroscope signals, we extracted a new set of features. Experimental results show that the new features we extracted perform well in common human activity recognition tasks using SVM, MLP, and CNN. On the other hand, we used a method where CNN automatically extracts features and performs recognition. Experimental results indicate that CNN's automatic feature extraction is more effective than manual extraction. The model proposed in this paper achieved a recognition rate of up to 98.53%, which is 2% higher than the current accuracy.
The long time data which contains the complete behavioral action is usually analyzed as a whole in human motion recognition methods of traditional radars, extracting global behavioral features for classification. This method does not work well in cases where the behavior contains multiple combinations of consecutive action. Based on millimeter wave radar, a solution in terms of time-frequency analysis and target detection of combination continuous motions is proposed to improve the recognition effect. First, a millimeter wave radar data acquisition platform is constructed to pre-process the collected echo signals; the projection method is then used to analyze the human movement echoes in time and frequency, and the FT-SSIM(Feature Transformation-structural similarity) algorithm is used to enhance the data; finally, a combination of YOLOv5 (You Only Look Once v5) and DeepSORT(Deep Simple Online and Realtime Tracking) is used to identify and count multiple combinations of consecutive action. Experimental results show that the recognition architecture proposed in this paper has good recognition effect on combined continuous action and average recognition accuracy of 97.5% after 300 iterations.
Deep learning technology has driven continuous advancements in the visual tracking field. In order to overcome various challenges, Siamese-based trackers and Attention-based trackers improve tracking performance by adding or deepening the network structure. However, they also increase training costs and model complexity. Lightweight trackers focus on achieving efficient and real-time tracking with limited computational resources, but these methods often face issues with lower accuracy and robustness in complex scenarios. Our approach Gradient Activation Siamese Network(Grad-Siam), adds fully connected layers to the backbone of SiamFC to obtain gradient activation information. By using the gradient activation strategy, we dynamically update the feature extraction process of the template and search branches. We trained on GOT-10k and validated on multiple standard test datasets. Experiments show that Grad-Siam balances accuracy and efficiency well, while also enhancing robustness in complex scenarios.
Attribute-Based Keyword Search (ABKS) technology has changed the way data is encrypted and retrieved, enabling data owners to securely store encrypted data in the cloud and allow retrieval only by users who comply with specific access policies. To prevent malicious tampering and protect user privacy, data owners need to revoke invalid users’ access to keyword ciphertexts in cloud storage. In addition, most ABKS schemes are vulnerable to keyword guessing attacks because any entity can generate keyword ciphertexts. This paper proposes a secure and efficient revocable attribute-based keyword search (SRABKS) scheme. First, SRABKS employs a revocation mechanism that enables the data owner to revoke invalid data users by updating the access policy with the help of the cloud server. In addition, we integrate the data owner’s private key into the keyword’s ciphertexts, thus effectively preventing unauthorized entities from generating invalid ciphertexts. Under the Decision Bilinear Diffie-Herman (DBDH) assumption, we give formal proofs of the indistinguishable security of keyword ciphertexts under adaptive chosen-keyword attack (KC-IND-CKA) and the indistinguishable security of keyword trapdoors under adaptive chosen-keyword attack (TD-IND-CKA). Finally, experiments show that the computational efficiency of SRABKS is very efficient.
Knowledge distillation for recommender models is a very challenging task because there can be millions of users and items, and The main focus is on determining the relative order of items based on user preferences, rather than predicting a label as in conventional classification models. Almost all the studies in the literature tackle these challenges by sampling items highly ranked by the teacher model and instructing the student model to mimic the teacher's behavior over the sampled items. However, such an approach may not work well when the gap between the teacher and the student is large, not to mention that learning from a small set of data often leads to overfitting. In this work, we propose a dual knowledge distillation approach that addresses these problems. The proposed knowledge distillation approach consists of two stages: 1) distillation from the teacher model to multiple assistant models of different structures, and 2) distillation from the assistant models to the target student model. We believe that when the assistants have different structures, they can be specialized in distinct areas of the teacher's knowledge space, thereby providing better guidance to the student's learning. We use the CiteULike dataset for our experiments and adopt hit ratio and normalized discounted cumulative gain as performance metrics. The results show that, compared to previous studies, incorporating multiple assistant models with different structures between the teacher and student models significantly enhances the recommendation performance of the student model. These improvements are consistent across all performance metrics used.
Automatic writing evaluation systems can provide language learners with instant feedback and suggestions on written texts, a useful tool that can benefit their writing abilities and skills. However, the development of such a system for Chinese university students of Portuguese remains an issue to be explored. The present study combines the Delphi method and the Analytic Hierarchy Process to establish an indicator system for the development of an automated writing evaluation system for Portuguese classes in higher education, with the weight of each metric being specified and analyzed. The system consists of five first-level indicators and thirteen second-level indicators. The primary indicators and their respective weights are "Content" (36.5%), "Mechanics" (21.6%), "Language use" (20.1%), "Vocabulary" (16.3%) and "Organization" (5.5%). With regard to the secondary indicators, the main concerns of the evaluators are "Relevance", "Readability" and "Spelling". The system of indicators defined in this research constitutes an important basis on which automated writing evaluation applications and software for texts written in Portuguese can be built, which is suitable for the reality in which the number of Chinese university students studying Portuguese is growing rapidly and they can improve their capabilities in Portuguese writing with the help of such applications and software.
Asymmetrically Clipped Optical Orthogonal Frequency Division Multiplexing (ACO-OFDM) stands as one of the most promising modulation techniques in high-speed optical communications. However, enhancing the error rate or spectral efficiency of conventional ACO-OFDM in optical systems remains a challenge. Existing enhanced ACO-OFDM demodulation techniques often suffer from either high computational complexity or poor performance due to limitations like DC-offsets and low frequency noise. To address this challenge, we propose a novel ACO-OFDM demodulation scheme based on Hartley transform and Pulse Amplitude Modulation (PAM), named "Improved Noise Cancellation ACO-OFDM (INC ACO-OFDM)". This scheme combines the advantages of both Diversity-Combined ACO-OFDM (DC ACO-OFDM) and Noise Cancellation ACO-OFDM (NC ACO-OFDM) without employing Hermitian symmetry. The proposed INC ACO-OFDM technique uses a Virtual Clean Windows (VCW) identification process, allowing the use of even subcarriers of ACO-OFDM to enhance the demodulation performance, even in the presence of DC offset. Simulation results conducted using MATLAB R2021a over an Additive White Gaussian Noise (AWGN) channel demonstrate that the proposed INC ACO-OFDM technique achieves similar Bit Error Rate (BER) performance as DC ACO-OFDM with a significant reduction in computational complexity, providing a 3 dB power gain compared to ACO-OFDM with similar spectral efficiency.
