Abstract
Content polluters on social media, such as spammers and fraudsters, increasingly employ "camouflage" tactics—interspersing malicious content with legitimate posts copied from normal users. This creates a significant challenge for detection systems, as account-level labels do not specify which individual posts are harmful. This paper introduces CCPDA (Camouflaged Content Polluter detection with Discriminant Analysis), a novel framework that utilizes sparse representation and discriminant analysis to filter out camouflage noise and pinpoint the specific malicious posts that reveal a polluter's true identity.
The Challenge of Camouflage
The openness of social media platforms is frequently exploited by content polluters. Unlike traditional spam, modern polluters blend malicious activity with a high volume of legitimate content to evade detection and build trust with human readers. This tactic presents two primary hurdles:
- Label Scarcity: While security systems can often identify a "polluter" account (e.g., via suspension), they lack labels for individual posts. It is unclear which posts are the actual "pollution" and which are merely "camouflage."
- Cognitive Inertia: By maintaining a high ratio of normal content, polluters make their malicious posts harder for both humans and automated systems to distinguish from background noise.
Methodology: The CCPDA Framework
CCPDA addresses these issues by learning to weight individual posts rather than treating every post from a known polluter as inherently malicious. The framework incorporates several key technical components:
- Weighting Vector ($c$): The model assigns a weight to each post, aiming to maximize the weight of actual polluting content while minimizing the weight of legitimate camouflage.
- Sparse Representation ($L_1$-norm): Recognizing that malicious content is typically a small fraction of a polluter's history, the model uses $L_1$ regularization to enforce sparsity, effectively searching for the "needle in the haystack."
- Group Exclusive Penalty ($L_{1,2}$-norm): This penalty ensures that for every known polluter account, the model identifies at least some discriminant posts, preventing it from ignoring difficult accounts during training.
- Alternating Optimization: The system alternates between updating the classification model and the post weights using Stochastic and Proximal Gradient Descent.
Experimental Results
The researchers evaluated CCPDA using a real-world Twitter dataset consisting of approximately 1.15 million posts from 94,535 unique users. Ground truth was established by tracking account suspensions over a three-year period.
CCPDA significantly outperformed standard baselines, including Support Vector Machines (SVM) and Gradient Boosted Decision Trees (GBDT). As shown in the performance comparison, CCPDA achieved a superior balance between precision and recall.
Key Insights
- The "Alike" Assumption: The model leverages the insight that while camouflage content is random and varied, malicious content from different polluters often shares similar templates or targets.
- Precision vs. Recall: While post-centric models (like $SVM_P$) can achieve higher raw recall, they suffer from extremely low precision. CCPDA maintains a high precision of 89.17%, ensuring that flagged accounts are highly likely to be actual polluters, while achieving a significantly higher F-score (45.96%) than its competitors.
Conclusion and Future Work
CCPDA demonstrates that treating the identification of polluting posts as a feature selection problem within the label space effectively "denoises" training data. Future research may improve these results by incorporating deeper linguistic analysis to understand the semantics of camouflage and combining content-based analysis with network-based behavioral dynamics.
Summary updated