Latest ArticlesTraditional Text-to-SQL tasks typically require large scale annotated data for model training, which incurs significant time and labor costs. Consequently, data scarcity remains a critical bottleneck, particularly in domain specific applications. While data augmentation based on large language model (LLM) has emerged as a promising solution to this issue, most existing methods heavily rely on the distribution of original datasets and often generate synthetic data that lack reliability and interpretability. Addressing these challenges, we propose SCoT-SQL, a novel chain-of-thought guided data synthesis framework. This method effectively integrates natural language instructions with symbolic knowledge, specifically detailed database schema and table content, to guide LLM, thereby significantly reducing semantic errors during the generation process. To guarantee the high quality of synthetic data, SCoT-SQL incorporates a rigorous filtering mechanism that leverages both model self consistency voting and verification based on execution. Additionally, the framework employs a self feedback mechanism driven by chain-of-thought reasoning to calibrate synthetic data. This process not only rectifies logical inconsistencies in SQL queries but also generates explicit step by step reasoning trajectories, enhancing the interpretability of the dataset. Comprehensive experiments on the KaggleDBQA and ScienceBenchmark datasets demonstrate the efficacy of the proposed method. SCoT-SQL achieves execution accuracy improvements of 6.5% and 3% respectively, significantly outperforming previous state of the art data synthesis baselines and proving its effectiveness in low resource and complex scenarios.
Retrieval-augmented generation (RAG) systems extend the capability of large language models by incorporating external knowledge databases, enabling more accurate and up-to-date responses. However, this integration also introduces a novel privacy vulnerability: mapping attacks (MAs), which aim to infer whether a specific private fragment is indexed in the retrieval database and to uncover its retrieval behavior. Such attacks exploit both embedding-level signals and retrieval dynamics, posing a serious threat to sensitive data protection in RAG pipelines. Despite the growing awareness of privacy risks in RAG systems, there is currently no dedicated defense mechanism specifically designed to mitigate mapping attacks. We introduce AMRP-SIP, a dual-randomization framework that concurrently protects both embeddings of the documents and retrieval traces, while preserving state-of-the-art utility. AMRP-SIP comprises three lightweight stages. First, a random orthogonal projection compresses each query and document into a low-dimensional latent space, hiding raw embeddings and reducing downstream noise. Second, adaptive differential privacy injects cluster-adaptive Gaussian noise, ensuring (ε, δ) fragment-level privacy. Third, a score-dropout layer introduces randomness by perturbing similarity scores with noise and probabilistically dropping a portion of the retrieved documents with probability p, thereby obfuscating the retrieval trajectory. Experiments on Wiki-40B, PubMed-256, and IP-DataSet demonstrate that AMRP-SIP reduces the AUC of membership inference attack (MIA) from 0.75 to 0.27, while maintaining competitive performance in downstream RAG tasks.
The diffusion models have made significant achievements in the field of image generation, but they still face some challenges such as high computational cost, long inference latency, and poor image generation quality during image generation. Based on previous research, we propose an efficient adaptive diffusion model inference method based on the warmup-aware and noise reuse methods, which can optimize the performance of the diffusion models according to input conditions and solve the problems of slow inference speed and poor image generation quality. Specifically, our warmup-aware method can dynamically adjust the number of warmup time steps in the initial stage of the diffusion process according to the semantic complexity of the input text prompts to ensure efficient initialization. Subsequently, our noise reuse method dynamically determines the number of time steps for noise reuse in the iterative refinement stage, and reduces unnecessary computational overhead. And we achieve 2.66 times acceleration on 4 NVIDIA V100 GPUs, and compared with DistriFusion, our inference speed increases by 12.8% and the image generation quality improves by 5.03 times. These improvements indicate that our proposed method has significant effects in balancing inference efficiency and image generation quality, and provides a scalable solution for real-time image generation applications.
The identification of cancer driver genes is an important research area in oncology research, holding significant value for a deeper understanding of the mechanisms of cancer development and achieving precision medicine. Currently, most research methods primarily rely on graph data to capture the characteristics of driver genes. However, the evolution process of cancer development involves many complex interactions among various biomolecules, making it difficult to fully reveal its essence using information from only single modality. Therefore, how to efficiently integrate multi-modal biological data to comprehensively identify driver genes has become a critical issue that needs to be addressed. To this end, we propose a cancer driver gene identification method, titled as TGMNN (text-graph modal neural network), based on a combination of text and graph modality. This method builds upon the graph modality by further incorporating text semantic information related to genes and employs an interactive attention mechanism to achieve collaborative fusion of multi-modal data, thereby reducing the differences between modalities. Compared with existing methods, TGMNN can capture the information interactions more effectively between modalities, significantly improving the accuracy of driver gene predictions. Experimental results show that AUC value of this method reaches 0.897, outperforming existing methods and high-lighting the model’s effectiveness.
In recent years, large vision-language models (VLM) have demonstrated remarkable zero-shot prediction capabilities in image classification tasks. Existing VLM-based classification approaches typically rely on a set of predefined category labels or external textual knowledge bases as semantic prior to align and match visual content with language. However, in real-world applications, such predefined labels often fail to cover the vast and dynamic concept space of the open world, thereby limiting the adaptability and generalization ability of these models. To address this challenge, a more open-ended task, generative image classification (GIC), is introduced, in which no prior semantic labels are provided, and the target categories reside in an open-ended, potentially unbounded semantic space, substantially increasing the classification difficulty. To tackle the complexity of GIC, a granularity-guided multi-model collaboration framework is proposed, designed to fully leverage the strengths of VLM in both image understanding and text generation. In this framework, prompt templates are carefully crafted to guide VLM in generating candidate labels at varying levels of semantic granularity. Subsequently, the candidate label space is expanded by a language model that associates semantically related categories based on the generated image descriptions at each granularity level. Finally, the most relevant label is identified through a semantic similarity retrieval process in the embedding space. The effectiveness and superiority of the proposed method for the GIC task are validated through extensive experiments on multiple fine-grained image classification benchmarks.
Existing image captioning methods often struggle to simultaneously capture global semantics and fine-grained details, leading to overly generic descriptions or missing key information. To address this issue, we propose EAFCC, short for expert-aware fusion with common-specific collaboration, a multi-feature image captioning model with mixture-of-experts fusion. The model takes grid features, region features, and segmentation features as parallel inputs. In the encoding stage, a pre-fusion module first performs implicit alignment of heterogeneous visual features and models cross-feature common information within a shared semantic space, while preserving the basis for feature-specific representations. A shared Transformer encoder is then employed to learn the common representations, whose consistency and discriminability are further enhanced by a contrastive learning constraint. Meanwhile, three independent encoding branches are used to model feature-specific representations, so as to retain fine-grained complementary information associated with different feature types. A gated routing mechanism inspired by mixture-of-experts is further introduced to dynamically weight and fuse the specific information. Finally, the enhanced common features and the fused specific features are collaboratively integrated, processed by a skip-connected encoder, and fed into a Transformer decoder to generate image captions. Experimental results on MS-COCO dataset show that EAFCC achieves competitive performance on the CIDEr-D metric and outperforms existing single-feature and dual-feature methods, demonstrating the effectiveness of common-specific collaborative modeling and dynamic multi-feature fusion for image captioning.
Deep learning technology has been increasingly empowering the practice of dermatological disease diagnosis, leading to a steady improvement in its diagnostic performance. However, such data-driven deep learning models are plagued by severe bias issues. If systems with these biases are deployed in clinical applications, they are bound to cause unfair treatment towards specific groups, such as underrepresented groups. Ensuring the fairness of decision-making has become a critical and urgent problem that needs to be addressed. Existing research has indicated that enhancing fairness often comes at the cost of reduced diagnostic performance. Therefore, to achieve a better trade-off between diagnostic performance and fairness, while simultaneously alleviating the unfair treatment of underrepresented groups, we propose an enhanced balanced incremental distillation network (EBID-Net). Specifically, with the aid of balanced memory, the model leverages representative demographic groups to assist underrepresented groups in knowledge learning, while being incrementally trained by integrating distributions across different groups. Additionally, we incorporate global information into the contextual attention mechanism to capture correlated interactions between features across different spatial locations, thereby obtaining more robust feature representations during incremental learning. Experimental results show that EBID-Net outperforms other methods in fairness criteria and in the trade-off between fairness and diagnostic performance.
Dynamic heterogeneous networks have become the mainstream paradigm for next-generation networks. The explosive integration of massive heterogeneous smart terminals imposes stringent requirements on encrypted data storage and privacy preservation within the cloud-edge collaborative architecture. Encrypted deduplication technology can identify and eliminate duplicate encrypted data, effectively reducing storage costs for cloud-edge environments while preserving data privacy, thus holding high practical value. However, existing encrypted deduplication schemes suffer from limitations such as vulnerability to duplicate faking attacks, high computational overhead, and heavy storage burdens on edge nodes, which fail to meet practical application demands. To address these issues, we propose SEDedup, a cross-user encrypted data deduplication scheme designed for cloud-edge storage scenarios. First, a server-aided encryption method suitable for cloud-edge storage is designed to achieve efficient and secure cross-user encrypted deduplication while effectively defending against duplicate faking attacks. Second, leveraging the temporal locality of data from heterogeneous terminals, an epoch-based duplicate data detection method is designed to achieve high-performance duplicate detection while minimizing the storage overhead on edge nodes. Finally, by incorporating generalized deduplication technology, lossless encrypted deduplication for similar heterogeneous terminal data is realized to maximize storage efficiency. Security analysis and performance evaluation demonstrate that SEDedup outperforms existing schemes in terms of performance while effectively preserving data privacy.
The memory encryption mechanism in confidential computing environments renders the internal state of confidential virtual machines (CVMs) invisible to the host. While ensuring internal data security, this poses severe technical challenges for traditional hypervisor-based disaster recovery and backup solutions. Meanwhile, existing solutions relying on TEE secure processors suffer from severe performance bottlenecks, which heavily restricts the deployment of cloud confidential computing clusters and degrades their security resilience. We present ResiCVM (resilient confidential virtual machine (VM)), a resilient disaster recovery and backup framework based on intra-VM privilege isolation. ResiCVM shifts the state extraction and cryptographic protection logic from the legacy TEE secure processor to a high-privilege domain within the VM, strictly excluding the guest operating system from the trusted computing base (TCB). Additionally, we design an end-to-end cryptographic authentication protocol to ensure secure communication between the VM and the security management center during the state backup process. We implement a prototype of ResiCVM on the AMD EPYC 9745 platform. Experimental evaluations show that ResiCVM achieves a backup encapsulation and storage throughput 62.7 times higher than that of TEE secure processor-based schemes. Crucially, during concurrent multi-VM maintenance tasks, ResiCVM maintains stable execution efficiency, whereas TEE-based solutions suffer proportional performance degradation. Requiring zero host-side modifications, ResiCVM significantly improves disaster recovery performance and provides robust technical support for the secure operations and security resilience of confidential computing clusters.
With the widespread deployment of IoT terminal devices in scenarios such as industrial control, intelligent transportation, smart homes, and smart healthcare, massive amounts of semi-structured logs are continuously generated at the edge. Existing methods either rely on log template IDs and shallow sequence models, making it difficult to model complex semantics, or rely on large-scale pre-trained models, which cannot meet the constraints of real-time performance and resource consumption of terminal devices. To address these issues, we propose a log anomaly detection technology for IoT terminal devices. First, a semantic enhancement detection model based on BERT and Llama is designed, employing regularization replacement and windowing modeling to suppress dynamic parameter noise. A semantic projection module is used to align the representation spaces of the encoder and decoder, and a three-stage fine-tuning method is used to improve the sequence-level anomaly detection capability. Second, a cloud-edge collaborative detection framework for terminal devices is constructed, proposing a two-stage confidence distillation and latency-sensitive dynamic routing strategy. This allows the small edge model to undertake most of the low-cost detection tasks, with only low-confidence samples uploaded to the large cloud model for verification. Our experiments on the public datasets HDFS, BGL, and Thunderbird-mini show that the proposed semantic detection model achieves F1 scores of 99.7%, 91.6%, and 96.6%, respectively. The cloud-edge collaborative framework maintains an overall F1 score of 95.5% while requiring only an average cloud routing ratio of 26.6%, which is equivalent to reducing the amount of large model calls by 73.4%.