New 'wrong-physics backdoors' found in neural PDE operators

Researchers have identified a new vulnerability in neural partial differential equation (PDE) operators, termed "wrong-physics backdoors." This attack exploits reusable solver archives by subtly altering inputs to trigger outputs that are physically plausible but incorrect for the intended physical parameters. The method, demonstrated across various PDE cases and model architectures like Fourier Neural Operators and DeepONet, achieves high success rates while maintaining low prediction errors. This highlights a critical validation gap, as generic solver-like behavior is insufficient without verifying the provenance of the physical parameters used. AI

IMPACT Exposes a critical validation gap in AI models used for scientific simulation, potentially impacting reliability in critical applications.

RANK_REASON Academic paper detailing a new security vulnerability in a specific type of AI model. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• DeepONet
• Fourier Neural Operators
• LSTM
• Neural PDE operators
• Transformer
• wrong-physics backdoors

• paper
• safety

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