arXiv:2605.26501v1 Announce Type: new 
Abstract: Large Vision-Language Models (LVLMs) have transformed multi-modal understanding, excelling in tasks like image captioning and visual question answering by integrating visual and textual inputs. However, their robustness against adversarial attacks, particularly those exploiting both modalities, remains underexplored, posing risks to critical applications like autonomous driving and content moderation. Existing attacks focus on single modalities or require impractical white-box access, limiting their real-world relevance. In this paper, we introduce Multi-Modal Adversarial Synergy, a groundbreaking framework that crafts universal, black-box multi-modal attacks against LVLMs. MMAS simultaneously generates a texture scale-constrained universal adversarial perturbation for images and a learnable prompt perturbation for text, optimized jointly using only model queries. The image perturbation leverages wavelet-based texture constraints to ensure imperceptibility and robustness across diverse visual inputs. The text perturbation, constrained by an L-norm in the embedding space, maintains semantic coherence while steering outputs toward a target. A novel cross-modal regularization term aligns the perturbations' gradient directions, enhancing their synergistic impact and transferability across tasks and models. Extensive experiments show the strong universal adversarial capabilities of our proposed attack with prevalent LVLMs.

كشفت الأبحاث الحديثة عن وجود ثغرات في نماذج الرؤية-اللغة الكبيرة (LVLM)، خاصةً تعرضها للهجمات العدائية متعددة الوسائط. يقدم البحث إطارًا يسمى التآزر العدائي متعدد الوسائط، الذي يولد اضطرابات عدائية عالمية لكل من الصور والنصوص، مما يثير القلق بشأن متانة نماذج LVLM في التطبيقات الحيوية مثل القيادة الذاتية وإدارة المحتوى.

Investigaciones recientes han revelado vulnerabilidades en los Modelos de Visión-Lenguaje de Gran Escala (LVLM), especialmente su susceptibilidad a ataques adversariales multimodales. El estudio introduce un marco denominado Sinergia Adversarial Multimodal, que genera perturbaciones adversariales universales tanto para imágenes como para texto, lo que plantea preocupaciones sobre la robustez de los LVLM en aplicaciones críticas como la conducción autónoma y la moderación de contenido.

Des recherches récentes ont révélé des vulnérabilités dans les grands modèles de vision-langage (LVLM), en particulier leur susceptibilité aux attaques adversariales multimodales. L'étude introduit un cadre appelé Synergie Adversariale Multimodale, qui génère des perturbations adversariales universelles pour les images et le texte, soulevant des inquiétudes quant à la robustesse des LVLM dans des applications critiques telles que la conduite autonome et la modération de contenu.

Recent research has unveiled vulnerabilities in Large Vision-Language Models (LVLMs), particularly their susceptibility to multi-modal adversarial attacks. The study introduces a framework called Multi-Modal Adversarial Synergy, which generates universal adversarial perturbations for both images and text, raising concerns about the robustness of LVLMs in critical applications like autonomous driving and content moderation.

Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization

arXiv:2406.05670v3 Announce Type: replace 
Abstract: Modern machine learning pipelines leverage large amounts of public data, making it infeasible to guarantee data quality and leaving models open to poisoning and backdoor attacks. Provably bounding model behavior under such attacks remains an open problem. In this work, we address this challenge by developing the first framework providing provable guarantees on the behavior of models trained with potentially manipulated data without modifying the model or learning algorithm. In particular, our framework certifies robustness against untargeted and targeted poisoning, as well as backdoor attacks, for bounded and unbounded manipulations of the training inputs and labels. Our method leverages convex relaxations to over-approximate the set of all possible parameter updates for a given poisoning threat model, allowing us to bound the set of all reachable parameters for any gradient-based learning algorithm. Given this set of parameters, we provide bounds on worst-case behavior, including model performance and backdoor success rate. We demonstrate our approach on multiple real-world datasets from applications including energy consumption, medical imaging, and autonomous driving.

تم تطوير إطار عمل جديد لتأكيد قوة نماذج التعلم الآلي ضد تسميم البيانات وهجمات الأبواب الخلفية، مما يعالج ثغرة كبيرة في أنظمة الذكاء الاصطناعي الحديثة التي تعتمد على مجموعات بيانات كبيرة. يتيح هذا الإطار ضمانات مثبتة حول سلوك النموذج دون تعديل النموذج الأساسي أو خوارزمية التعلم.

Se ha desarrollado un nuevo marco para certificar la robustez de los modelos de aprendizaje automático contra el envenenamiento de datos y los ataques de puerta trasera, abordando una vulnerabilidad significativa en los sistemas de IA modernos que dependen de grandes conjuntos de datos. Este marco permite garantías probables sobre el comportamiento del modelo sin alterar el modelo subyacente o el algoritmo de aprendizaje.

Un nouveau cadre a été développé pour certifier la robustesse des modèles d'apprentissage automatique contre le poisoning de données et les attaques par porte dérobée, abordant une vulnérabilité significative dans les systèmes d'IA modernes qui s'appuient sur de grands ensembles de données. Ce cadre permet des garanties prouvables sur le comportement des modèles sans modifier le modèle sous-jacent ou l'algorithme d'apprentissage.

A new framework has been developed to certify the robustness of machine learning models against data poisoning and backdoor attacks, addressing a significant vulnerability in modern AI systems that rely on large datasets. This framework allows for provable guarantees on model behavior without altering the underlying model or learning algorithm.

Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization

Was this article worth reading? Share it

Ready to build your own newsroom?