Abstract
This research investigates the privacy risks associated with aggregating seemingly innocuous social media content across multiple platforms. By analyzing data from Twitter, Instagram, and Foursquare, the study demonstrates that cross-platform data significantly enhances user de-anonymization. Furthermore, the authors introduce a novel "informativeness" metric that accurately predicts real-world location visits using only a small, curated subset of a user's timeline. The findings highlight a critical gap in user awareness regarding the cumulative privacy impact of individual posts.
Introduction
While many internet users recognize general online privacy risks, quantifying the specific risk of a single post remains a challenge. Is a tweet about the weather as revealing as one about a gym visit? This paper addresses the lack of tangible metrics for information disclosure by demonstrating how "innocent" text shared across different platforms can be aggregated to break anonymity and infer sensitive behavioral traits, such as physical location visits.
Methodology
The study employs two primary case studies using a mix of standard NLP techniques and a novel scoring framework:
User De-anonymization
User timelines are treated as large text documents and projected into a TF-IDF weighted vector space. A unigram language model is then used to calculate the probability that a specific user authored a set of anonymous posts based on their known cross-platform digital footprint.
Information Valuation and Location Inference
To predict visits to specific venue types (e.g., breweries, gyms), the researchers trained an AdaBoost classifier. A core contribution is the informativeness score, which combines:
- Relevance: Derived from Gini Importance in decision trees, indicating how useful specific words are for venue prediction.
- Novelty: A decay function that prioritizes new information by reducing the weight of frequently used words.
This score is utilized within an active learning framework, where the classifier is iteratively fed the most informative posts to evaluate profiling speed and accuracy.
Dataset
The researchers utilized a parallel corpus of 618 users with linked accounts across Twitter, Instagram, and Foursquare, totaling 1.1 million Tweets, 18,000 Instagram posts, and 99,000 Foursquare check-ins.
Key Insights: Venue Predictability
A significant finding is that the predictability of venue visits varies greatly depending on the location type. The authors categorize venues into three groups:
- Quick-to-learn: Locations like "Gyms" where a few relevant keywords are sufficient for high prediction accuracy.
- Slow-to-learn: Venues requiring a broader, more novel vocabulary over time to achieve accurate inference.
- Hard-to-learn: Locations that remain difficult to predict regardless of the amount of social media data analyzed.
Results and Performance
The proposed informativeness metric proved highly effective. In the de-anonymization task, the model achieved up to 77.96% accuracy in matching anonymous posts to users using only 20% of an average timeline.
In location inference, active selection of just 50 highly informative posts yielded an average F1-score of 44.19, significantly outperforming the random selection baseline of 15.36. In some cases, such as predicting "Brewery" visits, using only 50 curated posts (F1: 66.15) actually outperformed using the user's entire timeline (F1: 50.88), as the full timeline often introduces predictive noise.
Conclusion and Future Directions
The study underscores the vulnerability created by cross-platform consistency in writing styles and content. Future work aims to develop user-facing tools, such as browser extensions, to provide real-time feedback on the privacy risk of a post before it is shared. Additionally, the informativeness metric could be applied to optimize data acquisition in Real-Time Bidding (RTB) advertising pipelines, focusing on high-value data snippets rather than bulk collection.
Summary updated