Graph Neural Networks (GNNs) have achieved remarkable success in various applications, while still exhibiting high vulnerability to backdoor attacks when applied to node classification. Existing single-category attack methods typically rely on adaptive triggers that force victim nodes to be misclassified into a fixed target label, but they often neglect the inherent structural and feature priors associated with the target category. In this work, we propose a novel and effective backdoor attack framework Category-Specific Trigger Backdoor Attacks (CSTBA), employing category-specific information to generate more natural and unnoticeable triggers. Specifically, we introduce a Category-Specific Subgraph Triggers Pool (CS-STP) to capture representative patterns of the target category, along with a Match-and-Attach Strategy (MAS) to unnoticeably attach triggers to victim nodes, thereby ensuring that the modifications remain effective and unnoticeable within the graph. Extensive experiments on multiple benchmark datasets demonstrate that our method significantly enhances both the attack success rate (ASRs) and the unnoticeability of the attack compared with existing single-category approaches, highlighting the critical importance of category-aware trigger design in GNN backdoor attacks.

Concept Bottleneck Models (CBMs) decompose image classification into a process governed by interpretable, human-readable concepts. Recent advances in CBMs have used Large Language Models (LLMs) to generate candidate concepts. However, a critical question remains: What is the optimal number of concepts to use? Current concept banks suffer from redundancy or insufficient coverage. To address this issue, we introduce a dynamic, agent-based approach that adjusts the concept bank in response to environmental feedback, optimizing the number of concepts for sufficiency yet concise coverage. Moreover, we propose Conditional Concept Bottleneck Models (CoCoBMs) to overcome the limitations in traditional CBMs’ concept scoring mechanisms. It enhances the accuracy of assessing each concept’s contribution to classification tasks and feature an editable matrix that allows LLMs to correct concept scores that conflict with their internal knowledge. Our evaluations across 6 datasets show that our method not only improves classification accuracy by 6% but also enhances interpretability assessments by 30%.

Multimodal Large Language Models (MLLMs) have shown promise in visual-textual reasoning, with Multimodal Chain-of-Thought (MCoT) prompting significantly enhancing interpretability.However, existing MCoT methods rely on rationale-rich datasets and largely focus on inter-object reasoning, overlooking the intraobject understanding crucial for image classification.To address this gap, we propose WISE, a Weak-supervIsion-guided Step-bystep Explanation method that augments any image classification dataset with MCoTs by reformulating the concept-based representations from Concept Bottleneck Models (CBMs) into concise, interpretable reasoning chains under weak supervision.Experiments across ten datasets show that our generated MCoTs not only improve interpretability by 37% but also lead to gains in classification accuracy when used to fine-tune MLLMs 1 .Our work bridges concept-based interpretability and generative MCoT reasoning, providing a generalizable framework for enhancing MLLMs in fine-grained visual understanding.

The emergence of vision-language models has transformed medical AI, enabling unprecedented advances in diagnostic capability and clinical applications. However, progress in dermatology has lagged behind other medical domains due to the lack of standard image-text pairs. Existing dermatological datasets are limited in both scale and depth, offering only single-label annotations across a narrow range of diseases instead of rich textual descriptions, and lacking the crucial clinical context needed for real-world applications. To address these limitations, we present Derm1M, the first large-scale vision-language dataset for dermatology, comprising 1,029,761 image-text pairs. Built from diverse educational resources and structured around a standard ontology collaboratively developed by experts, Derm1M provides comprehensive coverage for over 390 skin conditions across four hierarchical levels and 130 clinical concepts with rich contextual information such as medical history, symptoms, and skin tone. To demonstrate Derm1M potential in advancing both AI research and clinical application, we pretrained a series of CLIP-like models, collectively called DermLIP, on this dataset. The DermLIP family significantly outperforms state-of-the-art foundation models on eight diverse datasets across multiple tasks, including zero-shot skin disease classification, clinical and artifacts concept identification, few-shot/full-shot learning, and cross-modal retrieval. Our dataset and code will be publicly available at https://github.com/SiyuanYan1/Derm1M upon acceptance.

Electric power resources are essential for the efficient and orderly development of society. Accurate power load forecasting is a key driver for the low-carbon upgrade of power systems. Traditional forecasting methods often struggle to capture long-term dependencies. Additionally, extracting complex nonlinear features from data remains a significant challenge, making it challenging to meet the accuracy demands of modern power systems. Besides, current deep learning-based forecasting methods cannot simulate multi-granularity power load data. To address these challenges, this paper presents a Generative Pre-trained Transformer model, GPT4PLTS, designed for power data simulation and fine-grained power load forecasting. The model leverages the Transformer architecture, incorporating the first six layers of the GPT decoder structure. It utilizes a multi-head attention mechanism to extract temporal features and includes a time alignment layer to maintain the sequence of time-series data, addressing both short-term and long-term dependencies. Extensive experiments are conducted on load observations from 2000 enterprises. The results demonstrate that GPT4PLTS achieves high accuracy in data simulation and forecasts across different time granularities, particularly excelling in short and medium-term predictions. Future research could focus on optimizing the model structure to enhance the model’s generalization ability.

Identifying sensitive text data remains a critical challenge for industrial data protection, as conventional methods lack flexibility and demand extensive manual labeling. While Large Language Models (LLMs) offer a promising avenue for automated sensitive data recognition (SDR), their susceptibility to “hallucination” often leads to misclassification. To overcome this, we introduce an LLM-based SDR framework powered by stepwise Chain-of-Thought (CoT) reasoning. Our approach employs a rigorously designed prompt template to enable LLMs to iteratively reassess and refine their output, leading to precise identification of sensitive text positions and highly fine-grained recognition results. Experiments on a real-world dataset from the electric sector demonstrate the superior performance of our method in domain-specific SDR tasks, highlighting its strong potential for robust and scalable application in complex industrial environments.

Hypervisor is a VMM (Virtual Machine Monitor) that creates and runs multiple VMs (Virtual Machines) through abstracting resources from a physical machine. Hypercall is a special and crucial call used in virtualized systems as it serves as a main communication channel between VMs and the hypervisor. However, hypercall attacks occur when an attacker manipulates the communication channel, and it could cause abnormal VM status, potentially leading to the abnormal resource allocation of the host OS (Operating System) and crash of VMs. Therefore, the virtualized system should execute abnormal VM status detection to identify potential abnormal behaviors to protect VMs and the host OS; however, existing works are either for reconstructing the hypervisor or hardware isolation, not for the VM status detection for abnormal hypercall.This study develops a hypercall-oriented abnormal VM status detection system called HypercallDetector based on the following three innovations: 1) we implement a hypercall tracing based on eBPF to obtain the hypercall-related running status (including CPU usage, memory usage, network traffic, etc.) of each VM; 2) we implement a window division technology to divide the VM status into multiple status windows of the same size, and appropriate window size with balanced detection precision (95.0%) and latency (within 8.8 ms) obtained by proposing the window regulator; and 3) we implement a CS-H algorithm (Compressing Sensing for Hypercall) to distinguish whether the VM status is abnormal. HypercallDetector shows higher precision and lower latency than its opponent and consumes only 8.6% CPU of single core and 0.3% memory usage when starting 240 VMs.

Control flow is the execution order of individual statements, instructions, or function calls within an imperative program. Malicious operation of control flow (e.g., tampering with normal function addresses) leads to severe consequences such as data leakage and system crash. Control Flow Integrity (CFI) is a defense restricting the execution order of program within Control Flow Graph (CFG). IndexHooks is an existing CFI solution designed against forward function calls tampering (including direct and indirect jump). This solution constructs a read-only linear function table that stores function addresses during compilation. Then, IndexHooks checks the table to make program jump to the correct target address during runtime. However, IndexHooks faces limitations in backtracking CFG construction, which can lead to excessive memory usage; the linear structure of the function table is vulnerable to brute force tampering. Addressing the limitations of IndexHooks, this study develops an obfuscated CFI solution called LeafHooks. LeafHooks is implemented during compilation by the LLVM compiler, which performs static analysis and instrumentation on the LLVM Intermediate Representation (IR) code of a program. We make the following three innovations: 1) we propose a speculation-free identification method for indirect function calls by linear traversing and analyzing codes to obtain legal function information (function address); 2) we save this information into a function table in the form of a Balanced Binary Tree (also known as AVL), enhancing the fuzzification of function addresses to defend against brute force; 3) we design a method to simulate control tamper attacks on ARM64 architecture to verify the ability of LeafHooks to protection. LeafHooks shows less overhead than state-of-the-art solutions and reduces 2.9% and 0.55% overhead on average using UnixBench and Phoronix, respectively.

Generalized category discovery (GCD) utilizes seen category knowledge to automatically discover new semantic categories that are not defined in the training phase. Nevertheless, there has been no research conducted on identifying new classes using medical images and disease categories, which is essential for understanding and diagnosing specific diseases. Moreover, existing methods still produce predictions that are biased towards seen categories since the model is mainly supervised by labeled seen categories, which in turn leads to sub-optimal clustering performance. In this paper, we propose a new neighbor-guided unbiased framework (NGUF) that leverages neighbor information to mitigate prediction bias to address the GCD problem in medical tasks. Specifically, we devise a neighbor-guided cross-pseudo-clustering strategy, which exploits the knowledge of the nearest-neighbor samples to adjust the model predictions thereby generating unbiased pseudo-clustering supervision. Then, based on the unbiased pseudo-clustering supervision, we use a view-invariant learning strategy to assign labels to all samples. In addition, we propose an adaptive weight learning strategy that dynamically determines the degree of adjustment of the predictions of different samples based on the distance density values. Finally, we further propose a cross-batch knowledge distillation module to utilize information from successive iterations to encourage training consistency. Extensive experiments on four medical image datasets show that NGUF is effective in mitigating the model's prediction bias and has superior performance to other state-of-the-art GCD algorithms. Our code will be released soon.

Knowing exact demographic attributes of users is crucial for human-computer interaction, intelligent marketing and automatic advertising. Ubiquitous sensor devices yield massive volumes of temporal data which hide a lot of valuable demographic information. In this paper, we bridge the gap between sensor data and demographic prediction to obtain real attributes of users from popular sensor devices: pedometer, which is widely used in mobile devices. We propose a novel model named Separated Embedding and Correlation Learning (SECL) for demographic prediction. Specifically, SECL first process the input data with a separated embedding layer to disentangle task-specific features for interference eliminating, and then capture the hidden correlations between different tasks via a correlation learning layer, finally the refined task-specific features are fed into a multi-task prediction layer to predict demographic attributes. Experimental results show impressive performance of our model on a real-world pedometer dataset, which is made publicly available on https://github.com/deepdeed/SECL.