[2025-11-25] 📚 Updates in #PINN

(1) <a href="https://researchtrend.ai/papers/2310.11789" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">PINNsFailureRegion Localization and Refinement through White-box AdversarialAttack
(2) PINNsFailureRegion Localization and Refinement through White-box AdversarialAttack

🔍 More at researchtrend.ai/communities/PINN

NatGVD achieves up to 53.04% natural adversarial evasion against GNN‑based and transformer vulnerability detectors, per a study submitted on 6 Oct 2025. https://getnews.me/natgvd-natural-adversarial-attack-on-graph-based-vulnerability-detection/ #graphneuralnetworks #adversarialattack

Research reveals a sparse gradient‑based adversarial attack can tweak just a few demand readings yet trigger voltage drops and overloads in restoration simulations. Read more: https://getnews.me/adversarial-attacks-on-load-forecasts-undermine-power-restoration/ #adversarialattack #gridresilience

A Zero Query Black-box Adversarial (ZQBA) attack adds feature maps from a source DNN to clean images, causing misclassification with a single query. The code is on GitHub. Read more: https://getnews.me/zero-query-black-box-attack-reduces-model-queries/ #zqba #adversarialattack

A new Proxy Targeted Attack (PTA) uses multiple source and target modality proxies to broaden adversarial success on multimodal AI models while staying hidden from detectors. https://getnews.me/new-proxy-attack-boosts-generalizability-stealth-of-multimodal-ai/ #multimodalai #adversarialattack

A task‑agnostic adversarial attack that distorts attention maps in vision foundation models lowers accuracy across classification, captioning and segmentation (Sept 2025). https://getnews.me/attention-attack-on-ai-foundation-models-disrupts-downstream-tasks/ #foundationmodels #adversarialattack

Research (AdvI2I) shows image perturbations can bypass Safe Latent Diffusion (SLD) safety filters, producing NSFW output. Paper on arXiv (doi:10.48550/arXiv.2410.71471). Read more: https://getnews.me/adversarial-image-attacks-bypass-safety-in-i2i-diffusion-models/ #adversarialattack #diffusion

Boosting adversarial robustness! New defense strategies explored. 🤔 [\#ICML] [\#AdversarialAttack] [\#DeepLearning] [\#AI]

Adversarial robustness is key!🛡️ New defenses explored in our latest paper. Check it out! 💡[#ICML] [#AdversarialAttack] [#Robustness] [#AIsecurity]

大型語言模型在真實世界的應用越來越廣，隨之而來的風險也逐漸浮現。尤其像 ChatGPT 這種經過大量安全對齊的模型，我們團隊投入很多資源確保其預設行為安全。但「adversarial attacks」或「jailbreak prompts」仍然可能讓模型產生不符預期的輸出。過去針對圖像的 adversarial attacks 研究很多，但文字這種離散資料的攻擊難度更高，因為缺乏直接的梯度訊號。這也是目前 LLM 應用落地的一大挑戰，如何在實用性與安全性之間取得平衡，是現階段很重要的議題。

#LLM #AdversarialAttack #Jailbreak #模型安全

https://l