Researchers have developed DeepInvert, a new semi-supervised attack that can recover original tokens from obfuscated language model embeddings with significantly higher accuracy than previous methods. This attack leverages both labeled shadow data and unlabeled target embeddings, demonstrating that current obfuscation defenses like ObfusLM, SentinelLMs, TextObfuscator, and DPNR offer less protection than assumed. Experiments show DeepInvert outperforms existing attacks on most defenses, highlighting a trade-off between utility and security, as defenses that retain utility also retain enough structure for inversion. AI
IMPACT This research highlights significant vulnerabilities in current language model obfuscation techniques, potentially impacting the security of sensitive data processed by cloud-based AI services.
RANK_REASON Academic paper detailing a new attack method against existing defenses. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
