[Paper Review] Enhancing Transparency and Control when Drawing Data-Driven Inferences about Individuals

Daizhuo Chen, Samuel P. Fraiberger+2 more|arXiv (Cornell University)|Jun 26, 2016

Privacy, Security, and Data Protection8 references4 citations

TL;DR

This paper proposes a cloaking mechanism that enables users to inhibit data-driven inferences about personal traits by selectively hiding (cloaking) their Facebook Likes, using evidence counterfactuals to identify the minimal set of Likes responsible for inferences. Results show that users can suppress inferences with minimal effort—averaging only 5–6 Likes cloaked—though firms can counter this by switching to models like Naive Bayes that exploit feature correlations, making cloaking significantly harder.

ABSTRACT

Recent studies have shown that information disclosed on social network sites (such as Facebook) can be used to predict personal characteristics with surprisingly high accuracy. In this paper we examine a method to give online users transparency into why certain inferences are made about them by statistical models, and control to inhibit those inferences by hiding ("cloaking") certain personal information from inference. We use this method to examine whether such transparency and control would be a reasonable goal by assessing how difficult it would be for users to actually inhibit inferences. Applying the method to data from a large collection of real users on Facebook, we show that a user must cloak only a small portion of her Facebook Likes in order to inhibit inferences about their personal characteristics. However, we also show that in response a firm could change its modeling of users to make cloaking more difficult.

Motivation & Objective

• To investigate whether users can achieve meaningful control over inferences drawn from their social media data.
• To assess the feasibility of transparency in data-driven inferences by identifying the minimal evidence required for a prediction.
• To evaluate whether users can inhibit inferences by cloaking only a small subset of their Likes.
• To examine how changes in modeling approaches (e.g., Naive Bayes vs. logistic regression) affect the difficulty of cloaking.
• To explore the implications of such control mechanisms for privacy, fairness, and user autonomy in online targeting systems.

Proposed method

• The method uses evidence counterfactuals to determine which input features (Likes) would, if removed, prevent a model from making a specific inference about a user’s personal trait.
• It applies this to a linear model where each Like is a binary feature with an associated coefficient, and the model score is the sum of these weighted features.
• Cloaking is simulated by iteratively removing Likes in order of their impact on the model’s prediction score, starting with the most influential.
• The approach is tested across multiple models: logistic regression (LR), logistic regression with 100 SVD components (LRSVD), and Naive Bayes (NB), to compare cloaking difficulty.
• The study uses real Facebook data from Kosinski et al. (2013) to train models predicting traits such as sexual orientation, gender, and religiosity.
• It measures 'cloakability' as the number of Likes a user must cloak to reduce their predicted score below a threshold, distinguishing between true positives and false positives.

Experimental results

Research questions

• RQ1 How many Likes must a user cloak to successfully inhibit a data-driven inference about a personal trait?
• RQ2 How does the difficulty of cloaking vary between users whose inferences are correct versus incorrect?
• RQ3 To what extent do different machine learning models (e.g., LR, NB) affect the effort required to cloak inferences?
• RQ4 Can firms counter user control by changing their modeling approach to make cloaking more difficult?
• RQ5 What is the trade-off between model performance and cloaking resistance in statistical inference systems?

Key findings

• On average, users need to cloak only 5.5 Likes to inhibit an inference when using logistic regression or LRSVD models, indicating low effort for effective cloaking.
• Cloaking is significantly easier for false-positive users (those incorrectly inferred to have a trait) than for true positives, with a statistically significant difference (p < 0.05).
• The Naive Bayes model increases cloaking difficulty substantially, requiring users to cloak an average of 50.6 Likes to inhibit an inference, compared to 5.5 for LR and LRSVD.
• The increased difficulty under Naive Bayes arises because it treats Likes as conditionally independent, leading to 'double counting' of correlated Likes and amplifying their influence.
• Firms can make cloaking more difficult without restricting user data by simply switching to a model like Naive Bayes that exploits feature correlations.
• The results suggest that while current models allow users to exert meaningful control with minimal effort, firms can undermine this control by altering their modeling strategy.