ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs
研究使用ModaLens方法通过图像交换审计来衡量医学视觉-语言模型在有无报告情况下的图像敏感性变化,揭示报告可用性降低了图像敏感性。
研究使用ModaLens方法通过图像交换审计来衡量医学视觉-语言模型在有无报告情况下的图像敏感性变化,揭示报告可用性降低了图像敏感性。
This study addresses the deformation of continuous fiber-reinforced thermoplastic composites during robotic 3D printing, which arises from residual stress relaxation, drying, crystallization, and thermal stresses. To tackle this challenge, a hybrid modeling approach integrating physical mechanisms with data-driven techniques is proposed. The mechanical behavior of the prepreg is captured using a Kelvin–Voigt viscoelastic constitutive model, while a stabilized neural ordinary differential equation (Neural ODE) framework is introduced to describe the coupled drying and crystallization kinetics. Validated through dynamic mechanical analysis (DMA), differential scanning calorimetry (DSC) experiments, and full-scale printing trials, the resulting model accurately reproduces real-world deformation phenomena and demonstrates strong generalization and robustness—even beyond the temperature ranges observed during training.
This work addresses a critical limitation in existing 3D point cloud adversarial attacks, which predominantly focus on geometric perturbations while neglecting the influence of topological structure on model robustness. To bridge this gap, the paper introduces the first topology-aware adversarial attack framework that treats topological features as an explicit attack dimension. The proposed method employs an end-to-end differentiable architecture leveraging differentiable persistent homology representations and persistence diagram embeddings to jointly optimize a composite objective comprising topological discrepancy loss, misclassification loss, and geometric imperceptibility constraints. By enabling gradient-guided perturbations that alter semantic interpretation without compromising geometric fidelity, the approach challenges the conventional assumption that geometric preservation entails semantic consistency. Extensive experiments demonstrate state-of-the-art performance, achieving up to 100% attack success rates against PointNet and DGCNN across ModelNet40, ShapeNet Part, and ScanObjectNN benchmarks, while significantly outperforming existing methods across multiple perceptual metrics.
This work addresses the limitation of existing approaches that rely on post-deployment user feedback to identify privacy issues, which prevents proactive mitigation prior to software release. To overcome this, the authors propose Pre-PI, a novel method that enables pre-release prediction of privacy concerns for the first time. Pre-PI achieves this by semantically aligning candidate features with existing ones, mapping historical privacy-related user reviews, and simulating user feedback to automatically generate privacy risk summaries. By shifting from reactive to proactive analysis, Pre-PI facilitates early privacy risk mitigation. Experimental evaluation on three real-world applications demonstrates that Pre-PI outperforms the state-of-the-art method Hark by identifying valid privacy concerns earlier and more comprehensively, significantly enhancing the ability to detect privacy issues before deployment.
This study addresses the absence of a unified safety evaluation benchmark for current large language models (LLMs), which hinders the quantification of deployment risks in critical applications. To bridge this gap, we propose the first standardized, cross-architecture safety evaluation framework that systematically assesses the vulnerability of five representative LLMs under six categories of adversarial attacks. Furthermore, we introduce a deployable multi-layered external defense mechanism. Experimental results reveal that existing models exhibit vulnerability rates ranging from 11.9% to 29.8%, whereas our defense framework achieves an average detection accuracy of 83% with only a 5% false positive rate. Notably, the findings demonstrate no direct correlation between a model’s general capabilities and its safety robustness, offering empirical evidence and a practical solution for safer LLM deployment.
研究使用ModaLens方法通过图像交换审计来衡量医学视觉-语言模型在有无报告情况下的图像敏感性变化,揭示报告可用性降低了图像敏感性。
This study addresses the deformation of continuous fiber-reinforced thermoplastic composites during robotic 3D printing, which arises from residual stress relaxation, drying, crystallization, and thermal stresses. To tackle this challenge, a hybrid modeling approach integrating physical mechanisms with data-driven techniques is proposed. The mechanical behavior of the prepreg is captured using a Kelvin–Voigt viscoelastic constitutive model, while a stabilized neural ordinary differential equation (Neural ODE) framework is introduced to describe the coupled drying and crystallization kinetics. Validated through dynamic mechanical analysis (DMA), differential scanning calorimetry (DSC) experiments, and full-scale printing trials, the resulting model accurately reproduces real-world deformation phenomena and demonstrates strong generalization and robustness—even beyond the temperature ranges observed during training.
This work addresses a critical limitation in existing 3D point cloud adversarial attacks, which predominantly focus on geometric perturbations while neglecting the influence of topological structure on model robustness. To bridge this gap, the paper introduces the first topology-aware adversarial attack framework that treats topological features as an explicit attack dimension. The proposed method employs an end-to-end differentiable architecture leveraging differentiable persistent homology representations and persistence diagram embeddings to jointly optimize a composite objective comprising topological discrepancy loss, misclassification loss, and geometric imperceptibility constraints. By enabling gradient-guided perturbations that alter semantic interpretation without compromising geometric fidelity, the approach challenges the conventional assumption that geometric preservation entails semantic consistency. Extensive experiments demonstrate state-of-the-art performance, achieving up to 100% attack success rates against PointNet and DGCNN across ModelNet40, ShapeNet Part, and ScanObjectNN benchmarks, while significantly outperforming existing methods across multiple perceptual metrics.
This work addresses the limitation of existing approaches that rely on post-deployment user feedback to identify privacy issues, which prevents proactive mitigation prior to software release. To overcome this, the authors propose Pre-PI, a novel method that enables pre-release prediction of privacy concerns for the first time. Pre-PI achieves this by semantically aligning candidate features with existing ones, mapping historical privacy-related user reviews, and simulating user feedback to automatically generate privacy risk summaries. By shifting from reactive to proactive analysis, Pre-PI facilitates early privacy risk mitigation. Experimental evaluation on three real-world applications demonstrates that Pre-PI outperforms the state-of-the-art method Hark by identifying valid privacy concerns earlier and more comprehensively, significantly enhancing the ability to detect privacy issues before deployment.
This study addresses the absence of a unified safety evaluation benchmark for current large language models (LLMs), which hinders the quantification of deployment risks in critical applications. To bridge this gap, we propose the first standardized, cross-architecture safety evaluation framework that systematically assesses the vulnerability of five representative LLMs under six categories of adversarial attacks. Furthermore, we introduce a deployable multi-layered external defense mechanism. Experimental results reveal that existing models exhibit vulnerability rates ranging from 11.9% to 29.8%, whereas our defense framework achieves an average detection accuracy of 83% with only a 5% false positive rate. Notably, the findings demonstrate no direct correlation between a model’s general capabilities and its safety robustness, offering empirical evidence and a practical solution for safer LLM deployment.