New Code-Poisoning Attack Leaks Private Data from ML Models

Researchers have developed a novel privacy attack called Code-Poisoning Property Inference Attack (CPPIA) that targets machine learning models trained on private data. This attack, designed for code hosting platforms like GitHub and coding agents such as Codex, allows malicious code providers to embed properties into secret samples during training. When users download this poisoned code and train models with their private data, the adversary can later query the trained model to leak sensitive information. CPPIA reportedly achieves 100% attack accuracy without degrading the model's performance, offering a computationally lightweight and effective method for privacy breaches. AI

IMPACT This attack highlights significant privacy risks in code-sharing platforms, potentially impacting the trust and security of ML model development.

RANK_REASON Academic paper detailing a new type of privacy attack on machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Code-Poisoning Property Inference Attack
• Codex
• GitHub
• machine learning
• Property Inference Attacks

• paper
• safety

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