New framework enhances DNN testing with latent space mutation

Researchers have developed Latte, a new black-box testing framework for deep neural networks designed to improve the identification of model weaknesses. Latte operates by mutating inputs within the network's latent space, generating test cases that are semantically similar to original inputs but diverse and capable of revealing errors. Evaluations on multiple datasets and models show that Latte enhances fault exposure and behavioral diversity compared to existing methods, while maintaining low semantic drift from the source seeds. AI

IMPACT This research could lead to more robust and secure AI systems by improving methods for identifying and mitigating model weaknesses.

RANK_REASON This is a research paper detailing a new method for testing deep neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Deep Neural Networks
• Latte
• VQ-VAE

• paper
• safety

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