Content-Based Spam Filtering on Video Sharing Social Networks

Antonio da Luz Jr., Eduardo Valle, Arnaldo de Albuquerque AraújoarXiv 2011 · arXiv

Abstract

This research investigates the detection of spam in video-sharing social networks by analyzing visual content. Moving beyond easily manipulated metadata, the study introduces a "bag of differences" approach that compares video responses to their parent videos. Results indicate that context-aware visual features significantly improve the identification of unrelated or promotional content, demonstrating that visual context is essential for effective spam filtering in dynamic social media environments.

Introduction

Spam on platforms like YouTube has evolved beyond simple "bad" content; it is primarily a problem of relevance. A legitimate video, such as a celebrity interview, becomes spam when posted as a "video response" to an unrelated topic, like a cooking tutorial.

Example of a potentially unrelated or promotional video frame used in the study.

Most existing spam filters rely on metadata—tags, titles, and user logs—which sophisticated spammers can easily game. This paper addresses the challenge of "content-based" filtering: using the actual pixels of a video to determine if it belongs in a specific thread. This task is inherently difficult due to the high-level semantics involved and the extreme visual diversity of legitimate user-generated content.

Methodology

The researchers developed a classification pipeline designed to capture both static and temporal visual information:

  1. Feature Extraction: The system utilizes two types of low-level features: SIFT (Scale-Invariant Feature Transform) for spatial detail in keyframes and STIP (Space-Time Interest Points) for motion and temporal patterns.
  2. Visual Dictionary (BoVF): Using a "Bag-of-Visual-Features" approach, thousands of local features are quantized into a 5,000-word visual vocabulary. Each video is then represented as a histogram of these visual words.
  3. Latent Semantic Analysis (LSA): To handle visual "synonymy" (where different visual words represent the same concept), Singular Value Decomposition (SVD) projects histograms into a compact "topic space."
  4. Context-Awareness (The "Bag of Differences"): The paper's primary contribution is the comparison of the response video to its parent. By subtracting the feature vector of the original video from the response, the model captures how the response deviates from the source topic.
  5. Classification: A Support Vector Machine (SVM) with a linear kernel classifies the resulting vectors as either "spam" or "legitimate."

Experimental Setup and Datasets

As no standard benchmark for video response spam existed, the authors curated a dataset of 8,182 videos from 84 "Most Responded" YouTube threads. Manual inspection labeled 3,420 videos as legitimate and 4,678 as spam. For the core experiments, a balanced subsample of 1,000 legitimate and 1,000 spam videos was evaluated using 5-fold cross-validation.

A frame from a legitimate video response, showing the typical organic nature of user content.

Key Insights and Results

The evaluation revealed several critical findings regarding the nature of video spam:

  • Context is Essential: The most significant finding was that even the weakest context-aware model outperformed the best context-blind model. Visual content alone is rarely sufficient to define spam; the relationship to the parent video is the deciding factor.
  • LSA and Generalization: Moving from raw visual word counts to a "topic space" via LSA helped the system generalize better, particularly when using SIFT features. This allowed the model to understand that different-looking objects might belong to the same semantic theme.
  • Visual Signatures of Spam: Interestingly, models could filter some spam even without context. This suggests that spam videos, such as advertisements or "trolling" content, often share distinct visual characteristics or production styles that differ from organic user responses.
  • Performance Metrics: The "Topic Differences" approach using SIFT features achieved the best results, significantly increasing the True Positive rate while maintaining a lower False Positive rate compared to context-blind versions.

Conclusion and Future Work

While the results prove that content-based analysis is feasible, the authors note that error rates remain too high for fully autonomous moderation. For a social network to automatically delete content, the False Positive rate must be near zero to avoid silencing legitimate users. Currently, the system is best suited as a "pre-filter" to flag suspicious videos for human review. Future iterations may achieve higher robustness by combining visual "bags of differences" with audio analysis, metadata, and social network statistics like user reputation.

Summary updated

BibTeX
@misc{jr2011content,
  author = {Antonio da Luz Jr. and Eduardo Valle and Arnaldo de Albuquerque Araújo},
  title = {{Content-Based Spam Filtering on Video Sharing Social Networks}},
  year = {2011},
  eprint = {1104.5284},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1104.5284},
}
Figure