FINDING · EVALUATION
Across three real-world text datasets and three large language models, the adaptive dynamic sampling scheme achieves efficiency and capacity comparable to existing PSS methods that require explicit model distributions — confirming that relaxing the distribution-access requirement does not degrade steganographic performance.
From 2026-bai-provable-secure-steganography — Provable Secure Steganography Based on Adaptive Dynamic Sampling · Abstract / Evaluation · 2026 · USENIX Security 2026
Implications
- Because the scheme matches existing PSS baselines on bits-per-token capacity while requiring only a seed API, it is the preferred option for circumvention deployments where only a third-party LLM API is available.
- Evaluation across multiple LLMs implies the technique is not architecture-specific; tool developers can switch underlying models without re-engineering the steganographic layer.
Tags
Extracted by claude-sonnet-4-6 — review before relying.