Tracking Illicit Drug Dealing and Abuse on Instagram Using Multimodal Analysis

Abstract

Illicit drug trade via social media sites, especially photo-oriented Instagram, has become a severe problem in recent years. As a result, tracking drug dealing and abuse on Instagram is of interest to law enforcement agencies and public health agencies. However, traditional approaches are based on manual search and browsing by trained domain experts, which suffers from the problem of poor scalability and reproducibility. In this article, we propose a novel approach to detecting drug abuse and dealing automatically by utilizing multimodal data on social media. This approach also enables us to identify drug-related posts and analyze the behavior patterns of drug-related user accounts. To better utilize multimodal data on social media, multimodal analysis methods including multi-task learning and decision-level fusion are employed in our framework. We collect three datasets using Instagram and web search engine for training and testing our models. Experiment results on expertly labeled data have demonstrated the effectiveness of our approach, as well as its scalability and reproducibility over labor-intensive conventional approaches.