This research introduces techniques to securely hide messages within LLM-generated text that can withstand attempts to corrupt the hidden information.

• Addresses critical vulnerabilities in existing LLM-based steganography methods
• Develops new algorithms resistant to "re-randomization attacks" where small text changes destroy hidden messages
• Demonstrates practical methods for creating robust covert communication channels
• Provides security guarantees against adversarial tampering

The findings advance digital security by enabling covert communication that remains intact even when text undergoes minor modifications, crucial for situations requiring discreet information exchange in potentially monitored environments.

Robust Steganography from Large Language Models