Human-defined creativity is highly abstract, posing a challenge for multimodal large language models (MLLMs) to comprehend and assess creativity that aligns with human judgments. The absence of an existing benchmark further exacerbates this dilemma. To this end, we propose CreBench, which consists of two key components: 1) an evaluation benchmark covering the multiple dimensions from creative idea to process to products; 2) CreMIT (Creativity Multimodal Instruction Tuning dataset), a multimodal creativity evaluation dataset, consisting of 2.2K diverse-sourced multimodal data, 79.2K human feedbacks and 4.7M multityped instructions. Specifically, to ensure MLLMs can handle diverse creativity-related queries, we prompt GPT to refine the human feedback to activate stronger creativity assessment capabilities. CreBench serves as a foundation for building MLLMs that understand human-aligned creativity. Based on the CreBench, we fine-tune open-source general MLLMs, resulting in CreExpert, a multimodal creativity evaluation expert model. Extensive experiments demonstrate that the proposed CreExpert models achieve significantly better alignment with human creativity evaluation compared to state-ofthe-art MLLMs, including the most advanced GPT-4V and Gemini-Pro-Vision.

ABSTRACT During autonomous driving, accurate detection of vehicles and pedestrians is important. Under foggy conditions, vehicle and pedestrian targets in visual sensor images often appear blurred, occluded, or with reduced contrast, which significantly hinders feature extraction. To address these challenges, in this paper, based on YOLO11, namely, DB‐YOLO, a vehicle and pedestrian detection algorithm for foggy scenes, is proposed. First, in accordance with the atmospheric scattering theory of imaging under foggy conditions, a dual‐backbone feature extraction structure is designed to implicitly model the degradation mechanism of foggy images, fuse high‐level and low‐level features, and gradually increase the receptive field, improving the representation ability of the model for foggy images. Second, a C3K2_iAFF module is introduced. This module employs an attention‐based data fusion approach to achieve more efficient feature fusion by suppressing irrelevant features that are polluted by haze while improving useful features that are related to the target. Additionally, an efficient upconvolution block (EUCB) upsampling method is used in the neck network to further suppress haze noise and improve key features. The experimental results reveal that DB‐YOLO achieved mAP50 values of 56.6% and 40% on two public foggy datasets, representing improvements of 8.1% and 2.5%, respectively, over the baseline model. The proposed approach has better detection performance than multiple state‐of‐the‐art (SOTA) models do. This study provides effective technical support for increasing the perceptual robustness of autonomous driving systems under adverse weather conditions.

ABSTRACT Insulator defect detection is critical for ensuring the safe operation of power systems. However, due to the small scale of insulator defects and their susceptibility to external factors such as weather conditions during the detection process, the detection rates are often suboptimal. To address these issues, we propose a new model, BMF‐SDFormer. A novel backbone module, BMFE, is developed. It overcomes the limitations of ResNet‐18. By integrating the custom‐designed C2F_PMFE module and convolutional layers, BMFE efficiently extracts multi‐scale features. This not only enhances the nonlinear representation ability and cross‐stage fusion within the model but also significantly reduces the computational complexity. Additionally, we introduce the SDFA network module, which utilises a dual‐domain attention mechanism to enhance feature extraction, improving the fusion of small target features and mitigating image degradation issues. Finally, the Focaler‐GIoU loss function is utilised to address the imbalanced distribution of difficult and simple samples, enabling precise detection of small defects on multiple insulators in complex environments, such as insulator strings. Experimental results demonstrate that BMF‐SDFormer achieves an mAP50 of 93.8%, an mAP50:95 of 62.6%, a precision of 92.7% and a recall of 90.8% on the IFDD test dataset. Furthermore, in comparison with other mainstream algorithms, the proposed BMF‐SDFormer demonstrates superior accuracy and efficiency in insulator defect detection tasks, highlighting its practical effectiveness for real‐world applications.

The behavior of the domain name is actually demonstrated by the behavior of the IP address. By observing the traffic behavior of their resolved IPs, the maliciousness of the domain names can be further divided. Deep packets inspection, reverse engineering and other approaches based on clustering technique in detecting malware using domain generation algorithms(DGA) are inefficient and with lots of false positives in large-scale networks. To address these challenges, this paper introduces a novel idea to identify DGA-based malware via behavior analysis. More specifically, four different types of traffic behaviors are focused, such as the rhythmic behavior, the cyclical behavior, the access stable behavior and the service diversity behavior. These behaviors are characterized, modeled and evaluated by using the metrics of the number of flows in each period of time, the access interval, the number of corresponding communication IPs in each period of time and the number of application types. NetFlow data of the resolved IP corresponding to the domain names collected from China Education Research Network backbone (CERNET) is applied to verify our proposal. Experimental results demonstrate that the application of IP address traffic behavior analysis can detect C& C channels of DGA-based malware regardless of its payload content.