New attack logic targets concept-level reasoning in neural networks

Researchers have introduced a novel method for attacking neural networks by focusing on logical inconsistencies in concept-level reasoning rather than just output misclassification. This approach, termed concept-based property attacks, aims to improve the robustness of models by falsifying human-interpretable concepts like "red and octagonal implies stop sign." The study demonstrates that these concept-level falsifications can generalize classical adversarial attacks and lead to more effective adversarial training, reducing semantic biases in models. AI

IMPACT Introduces a new method for evaluating and improving neural network robustness by targeting conceptual reasoning.

RANK_REASON Academic paper detailing a new methodology for attacking neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• C-XAI: A conceptual framework for designing XAI tools that support trust calibration
• DagsHub
• Gotit.pub
• Hugging Face
• Neural Networks
• Raik Dankworth
• ScienceCast

• paper
• safety

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