New ATLAS framework efficiently discovers adversarial input sets for AI models

Researchers have developed ATLAS (Adaptive Trust-Regions for Latent Adversarial Searches), a new query-based framework designed to efficiently discover sets of adversarial inputs that can cause black-box learning systems to fail. ATLAS frames attack generation as an active learning level set estimation problem, combining calibrated approximations with a local-global sampling architecture to identify regions of the input space containing adversarial examples. This method aims to build representative adversarial sets that accurately reflect a target model's robustness, outperforming existing query-based black-box attacks on standard and adversarially trained models like MNIST, CIFAR, and ImageNet. AI

IMPACT This research could lead to more robust AI systems by enabling better identification of vulnerabilities.

RANK_REASON The cluster contains a research paper detailing a new framework for adversarial attacks on machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Atlas Alone
• BayesOpt
• Cifar
• ImageNet
• MNIST database
• NES
• SignHunter

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →