Introduction
Social media challenges frequently blur the line between harmless entertainment and life-threatening behavior. The "Kiki Challenge"—inspired by Drake’s song "In My Feelings"—became a global phenomenon where participants exited moving vehicles to dance alongside them. This trend led to numerous traffic accidents, injuries, and fatalities. While previous research has focused on "dangerous selfies" or text-based trends, the analysis of video-based challenges has been hindered by a lack of annotated datasets. Researchers from the MIDAS Lab address this gap by introducing the MIDAS-KIKI dataset and a deep learning framework to automatically identify dangerous content.
The MIDAS-KIKI Dataset
To facilitate research in this area, the authors curated the MIDAS-KIKI dataset. Using 703 unique hashtags on Twitter, they collected 25,000 tweets, which were filtered down to 2,000 unique videos.
A critical challenge in this work was defining "danger" objectively. The researchers implemented the "ambulance test" heuristic: "After watching this video, does police and/or ambulance need to be informed?" This approach resulted in a high Cohen’s Kappa of 0.94, indicating strong inter-annotator agreement. The final dataset consists of 220 "dangerous" videos and 1,780 "non-dangerous" videos, reflecting the natural imbalance of social media content.
Methodology
The proposed system utilizes a transfer learning approach for video classification. The core of the model is a VGG-16 backbone, pre-trained on ImageNet, which serves as a robust feature extractor.
To process temporal data, the VGG-16 model is wrapped in a TimeDistributed layer in Keras, allowing it to apply the same visual feature extraction to every frame in a video sequence. These features are then flattened and passed through a series of dense layers (512 units each) and dropout layers (0.5 rate) to mitigate overfitting. The final output is a binary classification identifying the video as either "Safe Video" or "Unsafe Video."
Results and Discussion
The model achieved an overall accuracy of 87%. Given the context of content moderation, the high Precision (0.96) is particularly significant, as it ensures that flagged videos are highly likely to be genuinely dangerous, thereby reducing the burden of false positives on human moderators.
| Metric | Score |
|---|---|
| Accuracy | 0.87 |
| Precision | 0.96 |
| Recall | 0.90 |
| F1-Score | 0.93 |
The study also highlights the role of celebrity amplification in spreading these trends. For instance, high-profile figures performing the challenge in precarious locations can catalyze risky behavior among millions of followers, necessitating automated tools to manage the resulting surge in content.
Conclusion and Future Work
The MIDAS-KIKI dataset and the associated VGG-16 model provide a baseline for automated public safety moderation on social media. While the current model relies solely on visual data, future iterations could incorporate metadata—such as tweet text and location—to improve detection accuracy. Additionally, exploring the model's ability to generalize to other emerging challenges, such as the "Bird Box" or "Tide Pod" trends, remains a vital area for future research.
Summary updated