Contributions
- Large-scale characterization of fake-image spread during a real-world crisis: Identified 10,350 unique tweets containing 8 distinct fake images of Hurricane Sandy; 86% were retweets, showing low original-content production but massive amplification.
- Network analysis of visual misinformation propagation: Demonstrated that follower relationships account for only 11% of retweet edges—most spread occurs via trending topics and search queries, independent of social ties. Found hyper-concentrated amplification: top 30 users (0.3%) produced 90% of retweet volume.
- Effective classification of fake vs. real images: Built two-class classifiers (Naive Bayes, Decision Tree) distinguishing fake images from real Hurricane Sandy images with 97% accuracy (Decision Tree on tweet features), showing tweet-level content (length, hashtags, URLs, sentiment, pronouns, emoticons) far outperforms user-level features (followers, verification status, account age).
- Temporal and social analysis: Analyzed sudden virality spikes (e.g., 12-hour latency before peak spread) and constructed retweet networks to identify the few users driving propagation.
Method
Data collection: 24/7 monitoring of Twitter using Streaming API, querying real-time trending topics hourly. Filtered 1.78M tweets for keywords "sandy" or "hurricane" (Oct. 20–Nov. 1, 2012). Identified 10,350 fake-image tweets by cross-referencing public lists (Guardian, MSN, Mashable, TheWeek) that curated fake Hurricane Sandy images.
Characterization: Temporal analysis plotting hourly tweet volume; identified sudden peaks and reconstructed retweet/reply graphs during peak hours (Oct. 29, 21:00–22:00 UTC). Algorithm to compute overlap between retweet and follower graphs: for each edge in the retweet network, check if it exists in the follower network; computed intersection/total percentage.
Classification:
- Feature engineering: Two feature sets:
- User features (F1): #friends, #followers, follower-friend ratio, #times listed, has URL, verified, account age
- Tweet features (F2): length, #words, contains question/exclamation marks, #uppercase chars, sentiment words, pronouns, emoticons, #mentions, #hashtags, #URLs, retweet count
- Training: 5,767 fake-image tweets and 5,767 real-image tweets (balanced); 10-fold cross-validation with Naive Bayes and Decision Tree (J48).
Results
- Temporal: Fake URLs appeared Oct. 29, 06:00 UTC; spike in retweets 12 hours later; within one hour of spike, massive growth in retweeter count.
- Network: 10,779,122 edges in follower network; 10,508 edges in retweet network; only 1,215 overlapping edges = 11% overlap. Conclusion: crisis tweets spread via trending/search, not follower relationships.
- Amplification: CDF shows top 30 users (0.3%) responsible for 90% of retweets; remainder distributed across tail.
- Classification accuracy:
- Naive Bayes: F1 (user features) 56.32%, F2 (tweet features) 91.97%, F1+F2 91.52%
- Decision Tree: F1 53.24%, F2 97.65%, F1+F2 96.65%
- Conclusion: tweet features massively dominate; user-level reputation (followers, verification) nearly useless for distinguishing fake-image tweets.
Notes
This is an early work on image misinformation detection and an early proof that tweet-level linguistic and behavioral features outperform user reputation for identifying false visual content. The 97% accuracy is notable but should be contextualized: the dataset of 8 distinct images is small, and the high performance may reflect homogeneity in fake-image URLs (same URLs retweet repeatedly); future work confirms this limitation affects later image detection work. The 11% follower-graph overlap is a key contribution, showing that crisis information (both true and false) spreads primarily through public channels (trending, search) rather than social ties—a pattern distinct from political misinformation spread in normal times. The paper is limited by its focus on a single event and a small set of images, and does not address how the detection model generalizes to new images or events. However, it established important findings on temporal dynamics, network structure, and feature importance for image-based misinformation on social media during crisis.