Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GSDFuse: Capturing Cognitive Inconsistencies from Multi-Dimensional Weak Signals in Social Media Steganalysis

Kaibo Huang, Zipei Zhang, Yukun Wei, TianXin Zhang, Zhongliang Yang, Linna Zhou

TL;DR

This paper tackles the detection of covert linguistic steganography in social media dialogues by exploiting cognitive inconsistencies that arise from textual fragmentation and complex dialogue structures. It introduces GSDFuse, a holistic framework that combines hierarchical multi-modal feature representations, data augmentation for imbalanced learning, adaptive cross-modal evidence fusion, and discriminative embedding learning to detect subtle signals. Through experiments on multi-platform social media datasets, GSDFuse achieves state-of-the-art performance against classic and provably secure steganographic algorithms, especially under sparse embedding regimes. The work highlights the importance of integrating semantic, local topological, and global structural cues for robust steganalysis and points to future directions such as cross-platform generalization and resilience to adaptive steganography.

Abstract

The ubiquity of social media platforms facilitates malicious linguistic steganography, posing significant security risks. Steganalysis is profoundly hindered by the challenge of identifying subtle cognitive inconsistencies arising from textual fragmentation and complex dialogue structures, and the difficulty in achieving robust aggregation of multi-dimensional weak signals, especially given extreme steganographic sparsity and sophisticated steganography. These core detection difficulties are compounded by significant data imbalance. This paper introduces GSDFuse, a novel method designed to systematically overcome these obstacles. GSDFuse employs a holistic approach, synergistically integrating hierarchical multi-modal feature engineering to capture diverse signals, strategic data augmentation to address sparsity, adaptive evidence fusion to intelligently aggregate weak signals, and discriminative embedding learning to enhance sensitivity to subtle inconsistencies. Experiments on social media datasets demonstrate GSDFuse's state-of-the-art (SOTA) performance in identifying sophisticated steganography within complex dialogue environments. The source code for GSDFuse is available at https://github.com/NebulaEmmaZh/GSDFuse.

GSDFuse: Capturing Cognitive Inconsistencies from Multi-Dimensional Weak Signals in Social Media Steganalysis

TL;DR

This paper tackles the detection of covert linguistic steganography in social media dialogues by exploiting cognitive inconsistencies that arise from textual fragmentation and complex dialogue structures. It introduces GSDFuse, a holistic framework that combines hierarchical multi-modal feature representations, data augmentation for imbalanced learning, adaptive cross-modal evidence fusion, and discriminative embedding learning to detect subtle signals. Through experiments on multi-platform social media datasets, GSDFuse achieves state-of-the-art performance against classic and provably secure steganographic algorithms, especially under sparse embedding regimes. The work highlights the importance of integrating semantic, local topological, and global structural cues for robust steganalysis and points to future directions such as cross-platform generalization and resilience to adaptive steganography.

Abstract

The ubiquity of social media platforms facilitates malicious linguistic steganography, posing significant security risks. Steganalysis is profoundly hindered by the challenge of identifying subtle cognitive inconsistencies arising from textual fragmentation and complex dialogue structures, and the difficulty in achieving robust aggregation of multi-dimensional weak signals, especially given extreme steganographic sparsity and sophisticated steganography. These core detection difficulties are compounded by significant data imbalance. This paper introduces GSDFuse, a novel method designed to systematically overcome these obstacles. GSDFuse employs a holistic approach, synergistically integrating hierarchical multi-modal feature engineering to capture diverse signals, strategic data augmentation to address sparsity, adaptive evidence fusion to intelligently aggregate weak signals, and discriminative embedding learning to enhance sensitivity to subtle inconsistencies. Experiments on social media datasets demonstrate GSDFuse's state-of-the-art (SOTA) performance in identifying sophisticated steganography within complex dialogue environments. The source code for GSDFuse is available at https://github.com/NebulaEmmaZh/GSDFuse.

Paper Structure

This paper contains 42 sections, 11 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries and Methodology
  4. Key Challenges in Linguistic Steganalysis
  5. Problem Formulation and Task Definition
  6. Overall Model Architecture
  7. Detailed Components
  8. Hierarchical Feature Representation
  9. Semantic Node Profiling
  10. Topological Contextualization
  11. Data Augmentation for Imbalanced Learning
  12. Cross-Modal Feature Integration
  13. Optimization for Discriminative Embeddings
  14. Experiments
  15. Dataset Construction and Description
  16. ...and 27 more sections

Figures (3)

  • Figure 1: An illustrative example of a social media dialogue tree. A steganographic message (Steg, User C) is subtly embedded among benign messages (Cover). This scenario highlights key challenges such as textual fragmentation, steganographic sparsity, and complex message interactions that contribute to systemic distortions and latent multi-faceted imperceptibility in social media steganalysis.
  • Figure 2: Our proposed GSDFuse architecture. This multi-stage pipeline is designed to identify cognitive inconsistencies by effectively aggregating multi-dimensional weak signals in social media steganalysis, leveraging diverse feature representations and optimized learning strategies.
  • Figure 3: Average training time (s) per epoch as a function of SRS for the Reddit, Tweet, and Sina datasets.