International Journal of Intelligent Systems最新文献

Deep learning is widely used in many fields, but the emergence of adversarial examples threatens the application of deep learning. Various methods have been proposed to defend against adversarial attacks. However, existing defense methods either can only detect adversarial examples without restoring their original classes or merely focus on verifying the input category and attempting to recover the classes of adversarial examples while lacking awareness of whether the input has been perturbed. To develop defense approaches that simultaneously achieve both detection and correction capabilities, a heterogeneous model combinatorial defense framework (HMCDF) is proposed for adversarial attacks in this paper. In particular, we first summarize the fundamental operations, block structures, and compositional patterns that constitute the model, while analyzing how these factors influence both the functionality and robustness of the model. According to the differences in the structure of the models, the models can be divided into isomorphic models and heterogeneous models. Then, we combine heterogeneous models to construct a heterogeneous model defense framework. Within this framework, as long as a majority of models can detect adversarial examples and restore their original labels, the voting mechanism used in the framework can determine whether the input has been perturbed, ultimately outputting legitimate labels through collective decision-making. To validate the performance, we conduct extensive experiments on three public datasets: CIFAR-10, SVHN, and Mini-ImageNet. After sufficient analysis of the simulation results, we find that our proposed method outperforms the others for the detection of adversarial attacks generated by the considered attack methods and can recover the classes of the adversarial examples.

Contextual information parsing is one of the most important subtasks of conversational KBQA. However, existing methods often assume the independence of utterance and model them in isolation. In this paper, we propose a Dependency paRsing-Enhanced converSational queStion AnswerinG systEm, DRESSAGE, which can effectively model long-range semantic dependencies in the conversation history. This is a multitask neural semantic parsing model. The model can perform explicit dependency parsing for several history questions and the current question and enhance the entity recognition module and the question encoding module with the parsing tree. The performance of the DRESSAGE model is tested on the widely used CSQA dataset and gets SOTA in the overall effect, which proves the effectiveness of this model.

Facial expression recognition (FER) remains a challenging task in computer vision. Recent works have shown excellent performance in overall recognition accuracy, but its accuracy significantly decreases when recognizing similar expressions. This is due to interclass homogeneity and intraclass heterogeneity. To address these issues, we propose a novel dual-stage network called DUAL, inspired by contrastive learning. First, we increase the distance between negative samples while reducing the distance between positive ones. This is achieved by dynamically updating pairs of comparison samples. Second, we introduce a two-stage network architecture. The first stage uses two branches to extract image features and facial keypoint features. These branches interact to learn coarse-grained features through mutual guidance. The second stage focuses on fine-grained features using scale-specific residual blocks. This allows the model to identify facial regions that are critical for recognizing expressions. We conducted extensive experiments on multiple datasets. The results show that DUAL surpasses state-of-the-art models in items of performance. Additionally, the model shows high accuracy even in noisy conditions, highlighting its robustness.

Artificial intelligence (AI) systems are increasingly adopted in high-stakes domains such as healthcare and finance, so the demand for transparency and interpretability has grown substantially. EXplainable AI (XAI) methods have emerged to address this challenge, but individual techniques often offer limited, fragmented insights. This paper introduces Meta-explainers, a novel ensemble-based XAI framework that integrates multiple explanation types—specifically relevance-based and counterfactual methods—into unified, multifaceted and complementary meta-explanations. Inspired by meta-classification principles, our approach structures the explanation process into five stages: generation, grouping, evaluation, aggregation, and visualization. Each stage is designed to preserve the unique strengths of individual XAI techniques while enhancing their interpretability and coherence when combined. Experimental results on both image (MNIST) and tabular (Breast Cancer) datasets show that Meta-explainers consistently outperform individual and state-of-the-art ensemble explanation methods in terms of explanation quality, as measured by established metrics. This work paves the way toward more holistic and user-centered AI explainability with a flexible methodology that can be extended to incorporate additional explanation paradigms.

This paper introduces a cost-effective prompt optimization strategy for ancient Chinese word segmentation using large language models, aiming to mitigate the substantial computational resources and training expenses of fine-tuning. We developed two knowledge-enhanced frameworks, a General Knowledge Prompt framework and a Domain-Specific Knowledge Prompt framework, and evaluated their effectiveness across various ancient Chinese corpora using seven mainstream LLMs, including ERNIE Bot, Qwen, SparkDesk, DeepSeek, ChatGPT, Gemini, and Copilot. Our findings confirm that both prompt frameworks enhance the segmentation capability of LLMs to varying extents, with the Domain-Specific Knowledge Prompt framework yielding the most significant improvements. Notably, the DeepSeek model achieves 94.01% F1 score (94.24% precision, 93.79% recall) on the test set, while the Qwen model demonstrates a remarkable 15.73% increase in the F1 score with the Domain-Specific Knowledge Prompt framework. Our ablation studies indicate that the entries Rules and Examples are the most crucial to the success of prompt frameworks, effectively addressing the challenges of rule inconsistency and insufficient annotated data.

With the rise of neural networks and pre-trained models such as BERT, abstractive text summarization techniques have received widespread attention. Nevertheless, traditional extractive text summarization methods still hold substantial research value due to their low computational cost, interpretability, and robustness. In algorithms like TextRank and its variants, graph nodes are typically constructed based on surface-level lexical features. These graphs often fail to incorporate many contextual relationships, such as coreference relationships among nodes, resulting in fragmented representations of key concepts. For edge construction, a sliding window of size T is commonly used to connect word nodes within the window. However, these methods often fall short in modeling the rich contextual dependencies embedded in the document. Several recent studies have demonstrated that semantic graphs can effectively improve the accuracy of text summarization. In this paper, we construct a more interpretable semantic graph from syntax trees and propose a novel unsupervised algorithm based on the personalized PageRank algorithm for summary extraction. We utilize tree transformation methods to enrich word-level information for graph construction, define node-merging rules to reduce graph complexity, use coreference chains to merge coreferring entities across sentences for enriching contextual links, and introduce the concept of Meta Node sets to capture thematic relationships that are not fully represented by syntactic dependencies or coreference chains alone. By clustering semantically related words, Meta Nodes enhance the graph’s ability to reflect deeper contextual coherence across the document. Compared with previous TextRank-based methods, our improvement yields significant ROUGE score boosts on the CNN-DM dataset. While the method was developed and evaluated using English-language datasets, its underlying design is language agnostic and can be adapted to other languages with suitable linguistic tools.

Brain tumors account for approximately 2.5% of cancer-related deaths. Accurate classification of brain tumor types is essential for timely diagnosis and enhancing survival rates. Convolutional neural networks (CNNs) have demonstrated state-of-the-art performance in computer-aided diagnosis of brain tumors; however, the quality and availability of medical data significantly influence this process. Medical data must adhere to stringent privacy regulations, such as the General Data Protection Regulation (GDPR) in the European Union and the Health Insurance Portability and Accountability Act (HIPAA) in the United States. Federated learning (FL) enables the sharing of only model update parameters during collaborative training on locally stored data. However, these parameters may inadvertently enable reconstruction of the original data. Furthermore, medical data often exhibit nonindependent and nonidentically distributed (non-IID) characteristics, impeding model training performance. To address these challenges, this paper proposes a scheme that partitions confidential data into multiple segments during FL training, ensuring that only a subset exceeding a predefined threshold can reconstruct the data. The proposed scheme guarantees enhanced security, distributed control, and fault tolerance. In addition, this paper introduces a Conditional Mutual Information (CMI) regularizer to mitigate variability in model predictions. By minimizing the Kullback–Leibler (KL) divergence between local and global feature distributions, the CMI regularizer substantially enhances performance and convergence stability. Extensive experiments conducted on the Figshare dataset with varying α-values for data distributions validate the efficacy of the proposed model. Compared to FedAvg, FedProx, and FedDyn at α = 0.3, as well as the central model, the proposed model achieves a top-1 accuracy of 92.94% on the Figshare dataset, surpassing FedProx, FedAvg, and FedDyn by 2.42%, 2.82%, and 3.53%, respectively. Federated IID achieves performance comparable to that of the central model, further demonstrating its viability for practical applications.

Because of the copyright issues of hyperspectral images continue to rise, in this paper, we propose to use a neural network–based watermarking model to protect the copyright. By applying normalization-based attention module (NAM) to deep dispersed watermarking with synchronization and fusion (DWSF), a NDWSF model is proposed for robust hyperspectral image watermarking. It consists of encoding, decoding, discrimination, and attack modules. The encoding and decoding modules are used for embedding and extracting watermarks. Discrimination module is proposed for improving the quality of watermarked image. The discrimination module and the encoding module are in an adversarial relationship to motivate the encoder to generate watermarks with stronger invisibility. Attack module is employed between embedding and extraction to improve robustness against compression and noise and geometric attacks. In order to more effectively utilize image features for watermarking, a kind of hybrid attention mechanism is employed in embedding and extraction by adding NAM. Experimental results show that the loss convergence and stability in training is improved. The peak signal-to-noise ratio of the proposed method is 48.08 dB, higher than other methods about 2.5 dB. The bit error rate of the proposed method is less than 2.5% for various hybrid attacks, showing good robustness.

The pathological development of abnormal blood vessels results in neovascularization as a major vision-threatening condition of diabetic retinopathy. The main factor behind pathological vessel growth results from retinal capillary depletion of oxygen that causes abnormal vascular development patterns. Early detection of these fundus image abnormalities requires precision because it enables ophthalmologists to provide effective treatment and make proper diagnoses. A multiple-step image processing system treats this problem. A fusion-based contrast enhancement method begins the process of enhancing diabetic retinopathy fundus image brightness and contrast. After the initial process, the system applies detail weighted histogram equalization to the green channel for better structural detail visualization. In the second stage, the proposed online tiger-claw algorithm segments abnormal neovascularization from normal blood vessels. Next, the combination of fuzzy zone-based clustering with optimization and classifier thresholding performs local identification along with highlight generation for neovascularized areas. Neovascularization detection makes use of a YOLOv5 neural network in the third stage through feature extraction and classification operations. A refined segmentation process occurs with the application of multistage gray wolf optimization. The proposed algorithm underwent testing through its application to the public datasets STARE, DRIVE, MESSIDOR, and DIARETDB1. Experimental tests indicate that the neovascularization region marking performed with 98.19% sensitivity and 96.56% specificity while reaching 99.27% accuracy. The proposed approach demonstrates 97.03% accuracy and 98.94% sensitivity, together with 97.17% specificity in neovascularization detection.