Text-to-SQL technology converts natural language queries into SQL statements for database retrieval. Recent advances in large language models (LLMs) have improved Text-to-SQL performance, but generated SQL often contains semantic or syntax errors that degrade user experience and system stability. Existing SQL error detection methods are costly, lack interpretability, and do not support error labeling. To overcome these issues, we propose SQL-Checker a specialized model for Text-to-SQL error detection. We first analyze common error factors in Text-to-SQL, and we design a novel data synthesis framework based on these error factors. This framework simulates error factors to construct a basic error SQL data, and then using an error analysis template to distill high-quality SQL error analysis data from large-scale models. For complex errors, a self-guided iterative distillation strategy further enhances data quality. SQL-Checker is then trained on this distilled dataset. Additionally, we refine SQL error labeling and, integrate error label recognition into the detection task, enabling macro-level cause analysis. Experiments show SQL-Checker achieves state-of-the-art results on multiple error detection datasets. Incorporating SQL-Checker into the Text-to-SQL pipeline also improves execution accuracy.

Federated learning (FL) is a promising approach for privacy-preserving machine learning, enabling collaborative model training across distributed devices without sharing raw data. However, FL faces significant challenges due to the nonindependent and identically distributed (non-IID) nature of data across devices, leading to difficulties in model convergence and generalization. In this paper, we propose FedDiT, a novel federated learning framework that combines knowledge distillation with vision transformers. FedDiT introduces the Distilled Vision Transformer (DTViT) model on the client side, incorporating a distillation token to enhance local learning and knowledge transfer. This approach significantly improves the robustness and performance of FL in non-IID environments. We validated FedDiT through extensive experiments on public datasets, and the results show that it outperforms existing FL methods in both accuracy and smoother convergence. Additionally, FedDiT achieves higher throughput compared to standard transformers and knowledge distillation methods, making it more efficient for practical deployment in federated learning scenarios.

Secure multi-party computation (MPC) allows multiple parties to jointly compute a task while preserving the privacy of their individual inputs. In recent years, researchers have extensively studied various security protocols to improve the efficiency of MPC and reduce communication overhead. However, the efficiency of MPC in an$n$-party setting still remains a challenge. In this paper, we introduce WINK, a MPC framework designed for the semi-honest security model that accommodates an arbitrary number of participants. We design an$n$-party multiplication protocol in the$n$-out-of-$n$arithmetic secret sharing setting and, based on this, construct independent multiplication triples. We leverage a RLWE-based oblivious linear evaluation (OLE) technique to batch generate multiplication triples between any two parties, significantly reducing communication overhead. Furthermore, we design a constant-round n-party comparison protocol based on the properties of odd-sized rings and garbled circuits. This protocol completes in only three rounds of communication with the help of additional non-colluding semi-honest servers, marking a significant improvement over previous$n$-party comparison protocols. We prove that our framework achieves universally composable security. Moreover, experimental results demonstrate that our proposed comparison protocol outperforms existing comparison protocols in terms of efficiency.

Split Learning (SL) is a distributed privacy-preserving learning methodology designed to address the challenges associated with the deployment of large-scale deep neural networks on medical devices, while simultaneously safeguarding the privacy of medical data. However, both the forward and backward propagation of the model require communication between the medical devices and high-performance servers, resulting in significant communication overhead and high latency. In this paper, to reduce the communication overhead from the client to the server during forward propagation, we propose an autoencoder layer based on attention mechanisms and triple compression. To reduce the communication overhead from the server to the client during backward propagation, we proposed an average loss threshold algorithm to decrease the frequency of client updates. Compared to the original Split Learning algorithm, after incorporating the method proposed in this paper, the communication overhead during forward propagation decreased by an average of 93%, and during backward propagation, it decreased by an average of 96%. The total communication overhead decreased by an average of 95%. The model’s accuracy loss was between 0% and 1%. In terms of communication compression, compared to the SOTA SL–BSL, the overall communication overhead was reduced by an average of 92%.

The integration of Artificial Intelligence (AI) in the healthcare sector has marked significant advancements, and Federated Learning (FL) has further facilitated the amalgamation of Federated Medical Imaging. However, this integration has also sparked concerns regarding data privacy. Incorporating Differential Privacy (DP) into gradients effectively mitigates privacy leaks but at the cost of impacting model accuracy. Current research delves into DP within FL, with a focus on strategies for privacy budget allocation and noise addition. Nevertheless, the dynamic privacy requirements and resource optimization for actual medical applications are often overlooked, leading to resource wastage. This study introduces an innovative algorithm based on gradient component for adaptive noise scale optimization and privacy budget allocation, thereby enhancing privacy management while maintaining model accuracy. Our findings reveal that, compared to traditional DP techniques, our approach achieves an average accuracy improvement of 3.47% in RSNA-ICH Acc and an enhancement of up to 172.33% in Dice score for Prostate MRI Dice under various privacy budgets, demonstrating the substantial efficacy of our method in the domain of federated medical imaging.

Text-to-SQL is a subtask in semantic parsing that has seen rapid progress with the evolution of Large Language Models (LLMs). However, LLMs face challenges due to hallucination issues and a lack of domain-specific database knowledge(such as table schema and cell values). As a result, they can make errors in generating table names, columns, and matching values to the correct columns in SQL statements. This paper introduces a method of knowledge injection to enhance LLMs’ ability to understand schema contents by incorporating prior knowledge. This approach improves their performance in Text-to-SQL tasks. Experimental results show that pre-training LLMs on domain-specific database knowledge and fine-tuning them on downstream Text-to-SQL tasks significantly improves the Execution Match (EX) and Exact Match (EM) metrics across various models. This effectively reduces errors in generating column names and matching values to the columns. Furthermore, the knowledge-injected models can be applied to many downstream Text-to-SQL tasks, demonstrating the generalizability of the approach presented in this paper.

With the rapid proliferation of social networks, individuals now have greater access to news with increased speed. Simultaneously, there has been a heightened emphasis on detecting and mitigating the dissemination of fake news. One notable limitation of existing fake news detection models is their inability to effectively integrate multi-modal features, as they typically only establish connections between unimodal features, neglecting the potential synergies and complementarity among different modes. To address this issue, we introduce the Multi-modal Adaptive Co-attention fusion Contrastive learning Network (MACCN) for enhancing the detection of fake news by improving the fusion of textual and visual features. Our approach commences by employing distinct encoders to construct a high-level feature space for each modality. Subsequently, the Adaptive Co-attention Fusion Network is employed to establish strong correlations between textual and visual features, leading to a comprehensive representation. Ultimately, the model's performance is further enhanced through the application of contrastive learning, resulting in a more precise detection of fake news. We conducted an extensive series of experiments on three diverse datasets, and the results conclusively demonstrate that MACCN adeptly captures the interplay between multi-modal features, surpassing the performance of state-of-the-art methods.

Visual Question Answering (VQA) as a cutting-edge domain blending computer vision and natural language processing has garnered significant research momentum. Nev-ertheless, latest VQA frameworks which leverage CNN and RNN technologies to extract features and delve into the multimodal interplay between text and images, encounter difficulties in integrating local features with global dependencies. This integration challenge hinders the models' ability to precisely grasp the pivotal aspects of the answer. Furthermore, the dearth of training data poses a significant impediment to models' effective learning of multimodal information exchange and problem semantics. To address these issues, our model introduces a new method called CMACC (Cross-Modal Adversarial Contrastive Learning Based on Co-Attention). Through cross-modal adversarial contrastive learning, combined with common attention, image and text information are integrated. Adversar-ial learning aligns the latent feature distribution between text and images, and contrastive learning aligns multimodal sample features in the same context to enhance the adaptability of the model. In addition, advanced data augmentation techniques are integrated to further enhance the model's adaptability to different scenarios and problem types. We conducted experimental evaluations on three widely used VQA datasets (VQA v1.0, VQA v2.0, and COCO-QA) and the results showed that CMACC has significant improvements in accuracy and generalization performance compared to traditional methods.

Federated Learning (FL), a distributed learning paradigm optimizing communication costs and enhancing privacy by uploading gradients instead of raw data, now confronts security challenges. It is particularly vulnerable to Byzantine poisoning attacks and potential privacy breaches via inference attacks. While homomorphic encryption and secure multi-party computation have been employed to design robust FL mechanisms, these predominantly rely on Euclidean distance or median-based metrics and often fall short in comprehensively defending against advanced poisoning attacks, such as adaptive attacks. Addressing this issue, our study introduces “Split-Aggregation", a lightweight privacy-preserving FL solution capable of withstanding adaptive attacks. This method maintains a computational complexity ofO(dkN+k3) and a communication overhead ofO(dN), performing comparably to FedAvg whenk= 10. Here,drepresents the gradient dimension,Nthe number of users, andkthe rank chosen during random singular value decomposition. Additionally, we utilize adaptive weight coefficients to mitigate gradient descent issues in honest users caused by non-independent and identically distributed (Non-IID) data. The proposed method’s security and robustness are theoretically proven, with its complexity thoroughly analyzed. Experimental results demonstrate that atk= 10, this method surpasses the top-1 accuracy of current state-of-the-art robust privacy-preserving FL approaches. Moreover, opting for a smallerksignificantly boosts efficiency with only marginal compromises in accuracy.

Traffic and the movement of people are inextricably associated with the potential spread of COVID-19. In Intelligent Transportation System (ITS), Deep Learning (DL) traffic detection approaches driven by transportation big data have significant application values in monitoring, counting and classifying traffic vehicle information during the COVID-19 epidemic blockade, while DL COVID-19 medical diagnostic technology is also very important. However, due to concerns about data privacy and security, traditional data-centralized DL techniques that require uploading training data from multiple cameras or hospitals are no longer suitable. Federated Learning (FL) as a novel collaborative privacy-preserving DL paradigm could address this issue well. Nevertheless, in FL, most existing works train learning models with full-precision weights and communicate them over multiple iterations, which may incur massive additional communication costs and disclose the privacy implied in the trained local models. To tackle these issues, we first propose a novel multi-modal tensor ring decomposition TR-TSVD that not only achieves efficient data reduction but also keeps the correlations among multi-modes. Afterward, applying TR-TSVD to the training process of a convolutional neural network under the FL framework to achieve the goal of reducing communication overhead while ensuring model performance. Additionally, since the weight parameters are transmitted with the TR-TSVD format, attackers cannot infer the data privacy without knowing the specific restoration method. Besides, the additively homomorphic encryption is leveraged to further preserve model security. Extensive experimental results on MNIST, BIT-Vehicle and COVID-CT datasets show that the proposed approach could achieve a better performance.

Oblivious transfer plays a fundamental role in the area of secure distributed computation. In particular, this primitive is used to search items in decentralized databases. Using a variant of smooth projective hash previously presented by Zeng , we construct a practical framework fort-out-of-noblivious transfer in the plain model without any set-up assumption. It can be implemented under a variety of standard intractability assumptions, including the decisional Diffie-Hellman assumption, the decisionalN-th residuosity assumption, the decisional quadratic residuosity assumption, and the learning with error problem. It is computationally secure in the presence of covert adversaries and only requires four rounds of communication. Compared to existing practical protocols with fully-simulatable security against covert adversaries or malicious adversaries, our framework is generally more efficient.