Research map / Large language models / Instruction and fine-tuning
Hide and Seek in Embedding Space: Geometry-based Steganography and Detection in Large Language Models
Charles Westphal, Keivan Navaie, Fernando Rosas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace parity bucketing with embedding hyperplane bucketing (ASCII encoding) (ablation) | Ministral-8B LoRA · In-distribution steganography, TrojanStego-style data (T) | exact match rate (%): 24 -> 43 | Improved |
| replace parity bucketing with embedding bucketing (ASCII encoding) (ablation) | Llama-3.1-8B LoRA · In-distribution steganography, TrojanStego-style data (T) | exact match rate (%): 17 -> 30 | Improved |
| replace parity bucketing with embedding bucketing (ASCII encoding) (ablation) | Llama-3.1-8B LoRA · In-distribution steganography, Wiki prompts (W) | exact match rate (%): 38 -> 48 | Improved |
| replace parity bucketing with embedding bucketing (ASCII encoding) (ablation) | Ministral-8B LoRA · In-distribution steganography, Wiki prompts (W) | exact match rate (%): 36 -> 29 | Hurt |
| use embedding encoding and embedding bucketing instead of ASCII+parity (ablation) | Llama-3.3-70B LoRA · In-distribution steganography, Wiki prompts (W) | exact match rate (%): 9 -> 19 | Improved |
| use embedding bucketing with ASCII encoding instead of ASCII+parity (ablation) | Llama-3.3-70B LoRA · In-distribution steganography, Wiki prompts (W) | exact match rate (%): 9 -> 15 | Improved |
| use embedding encoding with parity bucketing instead of ASCII+parity (ablation) | Llama-3.3-70B LoRA · In-distribution steganography, Wiki prompts (W) | exact match rate (%): 9 -> 18 | Improved |
| use embedding+embedding instead of ASCII+parity (ablation) | Ministral-8B full fine-tune · In-distribution steganography, TrojanStego-style data (T) | exact match rate (%): 63 -> 57 | Hurt |
| average over 8B models: ASCII+Embedding vs Embedding+Embedding vs Embedding+Parity relative gain over baseline (ablation) | 8B models · In-distribution steganography | relative improvement (%): 12% / 33% / 25% | Improved |
| add mixed-motive instruction data (75/25) to Trojan-Stego fine-tuning (ablation) | Llama-3.1-8B LoRA ASCII+Embedding · Capability benchmarks (MMLU, ARC-C, HellaSwag, WinoGrande) | mean capability change (pp): -9.1 -> -2.7 | Improved |
| add mixed-motive instruction data to Trojan-Stego fine-tuning (ablation) | Ministral-8B full fine-tune ASCII+Parity · Capability benchmarks (MMLU, ARC-C, HellaSwag, WinoGrande) | mean capability change (pp): -31.8 -> -24.9 | Improved |
| probe final-layer hidden states of fine-tuned versus base models for bit detection | Llama/Ministral 8B stego models · OOD prompts, linear probe bit accuracy | bit accuracy (%): 60 -> 95--100 | Improved |
Closest papers by meaning
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.85)
- Better Language Model Inversion by Compactly Representing Next-Token Distributions (NeurIPS 2025 · similarity 0.85)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.85)
- Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks (NeurIPS 2025 · similarity 0.84)
- Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction--Concealment Tradeoff in MLLMs (NeurIPS 2026 · similarity 0.84)
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
