Research map / Large language models / Retrieval and knowledge
STLC-KG:A Social Text Steganalysis Method Combining Large-Scale Language Models and Common-Sense Knowledge Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STLC-KG, a social text steganalysis method that enriches input text with ConceptNet knowledge graph features and fine-tunes ChatGLM2-6B with a GAT branch and contrastive loss to detect linguistic steganography.
Problem
Detecting linguistic steganography in social media texts is hard because individual texts carry limited information, and LLM-generated stego text is highly fluent, so single-text steganalysis performs poorly.
Approach
Expand each text with a commonsense knowledge graph (ConceptNet) built into a knowledge feature graph encoded by a GAT, and combine it with LLM (ChatGLM2-6B) semantic features fine-tuned via prompt-based classification, with an added contrastive loss.
Key results
- GLM+KE (STLC-KG) achieves the highest accuracy on all six AC/ADG settings, averaging a 4.18% improvement over the best baseline SSM.
- On mixed-algorithm detection, GLM+KE reaches 80.09% accuracy on NEWS and averages a 2.87% gain over the best existing model.
- Using 45 knowledge items per mention was found to be the best setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · TWEET AC | accuracy: 75.02 -> 76.99 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · TWEET ADG | accuracy: 74.98 -> 76.57 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · MOVIE AC | accuracy: 82.70 -> 84.76 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · MOVIE ADG | accuracy: 81.36 -> 83.73 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · NEWS AC | accuracy: 83.87 -> 85.68 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · NEWS ADG | accuracy: 82.91 -> 84.49 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · TWEET mixed steganography | accuracy: 70.02 -> 70.69 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · MOVIE mixed steganography | accuracy: 73.38 -> 74.76 | Helped |
| add ConceptNet knowledge enhancement to GLM (ablation) | GLM steganalysis · NEWS mixed steganography | accuracy: 78.65 -> 80.09 | Helped |
| use Freeze rather than LoRA or P-Tuning V2 (ablation) | ChatGLM2-6B steganalysis · TWEET | F1 score: 70.07; 68.63; 68.99 | Helped |
| use P-Tuning V2 rather than Freeze or LoRA (ablation) | ChatGLM2-6B steganalysis · MOVIE and NEWS | F1 score: 72.30; 70.44; 73.50; 76.61; 75.82; 78.75 | Helped |
Surprising findings
- Freeze fine-tuning worked best on the short TWEET dataset while P-Tuning v2 was best on the longer MOVIE and NEWS datasets.
Limitations
- Learning rate value left unspecified in the paper
- Improvements over baselines are modest on some settings (e.g., 2.51% on NEWS ADG)
- Knowledge base limited to ConceptNet word-level relations and excludes some relation types
Experimental setup
Datasets: TWEET MOVIE NEWS ConceptNet
Benchmarks: TWEET MOVIE NEWS
Metrics: accuracy precision recall F1
Baselines: FCN RBC BERT SeSy SSM GLM
Builds on: ChatGLM2-6B ConceptNet Graph Attention Network RoPE P-Tuning v2
Closest papers by meaning
- Promising Multi-Granularity Linguistic Steganography by Jointing Syntactic and Lexical Manipulations (AAAI 2025 · similarity 0.84)
- STEAD: Robust Provably Secure Linguistic Steganography with Diffusion Language Model (NeurIPS 2025 · similarity 0.84)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.83)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.82)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.82)
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.82)
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.