Statistically undetectable backdoors planted in deep neural networks

Researchers have demonstrated a method for embedding statistically undetectable backdoors into deep neural networks. These backdoors are designed to be imperceptible even with full model access, meaning the backdoored model is nearly identical to an honestly trained one in terms of total variation distance. The backdoor enables the generation of invariance-based adversarial examples, which are otherwise provably impossible to create in polynomial time without the backdoor. This work highlights a significant power imbalance between those who train models and those who use them. AI

IMPACT Reveals a fundamental security vulnerability in deep learning models, potentially impacting trust and security in AI systems.

RANK_REASON Academic paper detailing a new security vulnerability in machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• Deep Neural Networks
• machine learning

• paper
• safety

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