arXiv:2501.01015v2 Announce Type: replace 
Abstract: Deep neural networks are vulnerable to adversarial examples that exhibit transferability across various models. Numerous approaches are proposed to enhance the transferability of adversarial examples, including advanced optimization, data augmentation, and model modifications. However, these methods still show limited transferability, particularly in cross-architecture scenarios, such as from CNN to ViT. To achieve high transferability, we propose a technique termed Spatial Adversarial Alignment (SAA), which employs an alignment loss and leverages a witness model to fine-tune the surrogate model. Specifically, SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. First, we minimize the divergences of features between the two models in both global and local regions, facilitating spatial alignment. Second, we introduce a self-adversarial strategy that leverages adversarial examples to impose further constraints, aligning features from an adversarial perspective. Through this alignment, the surrogate model is trained to concentrate on the common features extracted by the witness model. This facilitates adversarial attacks on these shared features, thereby yielding perturbations that exhibit enhanced transferability. Extensive experiments on various architectures on ImageNet show that aligned surrogate models based on SAA can provide higher transferable adversarial examples, especially in cross-architecture attacks.

تتناول دراسة حديثة التحديات التي تواجه الأمثلة العدائية في الشبكات العصبية العميقة، وخاصة محدودية قابليتها للنقل عبر هياكل مختلفة مثل CNN وViT. تسلط الأبحاث الضوء على استراتيجيات متعددة، مثل تحسين البيانات وزيادة البيانات، تهدف إلى تحسين هذه القابلية للنقل. فهم هذه الديناميكيات أمر بالغ الأهمية لأنه يمكن أن يؤدي إلى أنظمة ذكاء اصطناعي أكثر قوة، مما يعزز الأمان في التطبيقات التي يتم فيها نشر التعلم العميق.

Un estudio reciente discute los desafíos de los ejemplos adversariales en las redes neuronales profundas, particularmente su limitada transferibilidad entre diferentes arquitecturas como CNN y ViT. La investigación destaca varias estrategias, como la optimización y la augmentación de datos, destinadas a mejorar esta transferibilidad. Comprender estas dinámicas es crucial, ya que puede llevar a sistemas de IA más robustos, mejorando la seguridad en aplicaciones donde se utiliza el aprendizaje profundo.

Une étude récente aborde les défis des exemples adversariaux dans les réseaux de neurones profonds, en particulier leur transférabilité limitée entre différentes architectures comme les CNN et les ViT. La recherche met en lumière diverses stratégies, telles que l'optimisation et l'augmentation des données, visant à améliorer cette transférabilité. Comprendre ces dynamiques est crucial car cela peut conduire à des systèmes d'IA plus robustes, renforçant ainsi la sécurité dans les applications où l'apprentissage profond est déployé.

A recent study discusses the challenges of adversarial examples in deep neural networks, particularly their limited transferability across different architectures like CNNs and ViTs. The research highlights various strategies, such as optimization and data augmentation, aimed at improving this transferability. Understanding these dynamics is crucial as it can lead to more robust AI systems, ultimately enhancing security in applications where deep learning is deployed.

Boosting Adversarial Transferability with Spatial Adversarial Alignment

arXiv:2601.08100v1 Announce Type: new 
Abstract: Understanding the generalization behavior of deep neural networks remains a fundamental challenge in modern statistical learning theory. Among existing approaches, PAC-Bayesian norm-based bounds have demonstrated particular promise due to their data-dependent nature and their ability to capture algorithmic and geometric properties of learned models. However, most existing results rely on isotropic Gaussian posteriors, heavy use of spectral-norm concentration for weight perturbations, and largely architecture-agnostic analyses, which together limit both the tightness and practical relevance of the resulting bounds. To address these limitations, in this work, we propose a unified framework for PAC-Bayesian norm-based generalization by reformulating the derivation of generalization bounds as a stochastic optimization problem over anisotropic Gaussian posteriors. The key to our approach is a sensitivity matrix that quantifies the network outputs with respect to structured weight perturbations, enabling the explicit incorporation of heterogeneous parameter sensitivities and architectural structures. By imposing different structural assumptions on this sensitivity matrix, we derive a family of generalization bounds that recover several existing PAC-Bayesian results as special cases, while yielding bounds that are comparable to or tighter than state-of-the-art approaches. Such a unified framework provides a principled and flexible way for geometry-/structure-aware and interpretable generalization analysis in deep learning.

تشير دراسة جديدة إلى إطار موحد لـ PAC-Bayes للحدود العامة المعتمدة على المعايير، حيث تتناول التحديات المتعلقة بفهم سلوك التعميم في الشبكات العصبية العميقة. تعيد هذه الدراسة صياغة اشتقاق هذه الحدود كمشكلة تحسين عشوائي على توزيعات غاوسية غير متساوية، بهدف تعزيز الصلة العملية للنتائج.

Un nuevo estudio propone un marco PAC-Bayes unificado para los límites de generalización basados en normas, abordando los desafíos de comprender el comportamiento de generalización de las redes neuronales profundas. La investigación reformula la derivación de estos límites como un problema de optimización estocástica sobre posteriors gaussianos anisotrópicos, con el objetivo de mejorar la relevancia práctica de los resultados.

Une nouvelle étude propose un cadre PAC-Bayésien unifié pour les bornes de généralisation basées sur les normes, abordant les défis de la compréhension du comportement de généralisation des réseaux de neurones profonds. La recherche reformule la dérivation de ces bornes comme un problème d'optimisation stochastique sur des postérieurs gaussiens anisotropes, visant à améliorer la pertinence pratique des résultats.

A new study proposes a unified PAC-Bayesian framework for norm-based generalization bounds, addressing the challenges of understanding deep neural networks' generalization behavior. The research reformulates the derivation of these bounds as a stochastic optimization problem over anisotropic Gaussian posteriors, aiming to enhance the practical relevance of the results.

Towards A Unified PAC-Bayesian Framework for Norm-based Generalization Bounds

arXiv:2601.07944v1 Announce Type: new 
Abstract: Since the turn of the century, approximate Bayesian inference has steadily evolved as new computational techniques have been incorporated to handle increasingly complex and large-scale predictive problems. The recent success of deep neural networks and foundation models has now given rise to a new paradigm in statistical modeling, in which Bayesian inference can be amortized through large-scale learned predictors. In amortized inference, substantial computation is invested upfront to train a neural network that can subsequently produce approximate posterior or predictions at negligible marginal cost across a wide range of tasks. At deployment, amortized inference offers substantial computational savings compared with traditional Bayesian procedures, which generally require repeated likelihood evaluations or Monte Carlo simulations for predictions for each new dataset.
  Despite the growing popularity of amortized inference, its statistical interpretation and its role within Bayesian inference remain poorly understood. This paper presents statistical perspectives on the working principles of several major neural architectures, including feedforward networks, Deep Sets, and Transformers, and examines how these architectures naturally support amortized Bayesian inference. We discuss how these models perform structured approximation and probabilistic reasoning in ways that yield controlled generalization error across a wide range of deployment scenarios, and how these properties can be harnessed for Bayesian computation. Through simulation studies, we evaluate the accuracy, robustness, and uncertainty quantification of amortized inference under varying signal-to-noise ratios and distributional shifts, highlighting both its strengths and its limitations.

أظهرت دراسة حديثة فعالية الاستدلال المدعوم في الإحصاءات البايزية، خاصة تحت تغيرات نسبة الإشارة إلى الضوضاء وتحولات التوزيع. تستخدم هذه الطريقة الشبكات العصبية العميقة لتبسيط عملية الاستدلال، مما يسمح بتوفير كبير في الحوسبة مقارنة بالأساليب البايزية التقليدية التي تتطلب تقييمات واسعة لاحتمالية.

Un estudio reciente ha evaluado la efectividad de la inferencia amortizada en estadísticas bayesianas, particularmente bajo variaciones en la relación señal-ruido y cambios en la distribución. Este método aprovecha redes neuronales profundas para agilizar el proceso de inferencia, permitiendo ahorros computacionales significativos en comparación con los enfoques bayesianos tradicionales que requieren evaluaciones extensas de verosimilitud.

Une étude récente a évalué l'efficacité de l'inférence amortie dans les statistiques bayésiennes, en particulier sous des variations de rapport signal-bruit et des changements de distribution. Cette méthode utilise des réseaux de neurones profonds pour rationaliser le processus d'inférence, permettant des économies computationnelles significatives par rapport aux approches bayésiennes traditionnelles qui nécessitent des évaluations de vraisemblance étendues.

A recent study has assessed the effectiveness of amortized inference in Bayesian statistics, particularly under varying signal-to-noise ratios and distribution shifts. This method leverages deep neural networks to streamline the inference process, allowing for significant computational savings compared to traditional Bayesian approaches that require extensive likelihood evaluations.

A Statistical Assessment of Amortized Inference Under Signal-to-Noise Variation and Distribution Shift

One More Thing in AI – Your Shortcut to AI Mastery

Boosting Adversarial Transferability with Spatial Adversarial Alignment

Was this article worth reading? Share it

One More Thing in AI

Humanize AI

LucidQuery AI

Attentive AI

Https

Blunge

Ready to build your own newsroom?