Multi-Dimensional Behavioral Signature Analysis for Video Bullet Comment Steganography Detection

The proliferation of bullet comment systems on video-sharing platforms has created novel opportunities for covert communication. Recent research has demonstrated multiple bullet comment-based steganographic paradigms: time modulation, time attribute shifting, live broadcast game-based channels, and generative stegotext via frame comments. This paper presents Multi-Dimensional Behavioral Signature Analysis (MDBSA), a unified detection framework integrating seven behavioral dimensions to systematically characterize anomalies from four documented bullet comment steganographic paradigms. We construct a synthetic benchmark spanning six video categories with controlled steganographic injection; extract features capturing temporal, spatial, content, and structural patterns; and evaluate it using GroupKFold cross-validation to prevent data leakage from overlapping sliding windows. On this synthetic benchmark, MDBSA features combined with Gradient Boosting achieve 96.9% accuracy (F1 = 96.8%, AUC = 0.996), compared with 65.6% for logistic regression on the same features. Per-paradigm detection rates on the benchmark range from 98.4% to 100.0%.