Suchergebnisse

Intro -- Acknowledgments -- Introduction -- Chapter Overview -- Terminology and Definitions -- Attributes -- Privacy Settings of Attributes -- Risk Sources -- Data Inference -- Threat -- Privacy Harm -- Privacy Risks -- Privacy Risk Analysis -- Dimensions of Privacy Scoring in OSNs -- Type of Data -- Assumptions About the User -- Privacy Settings -- Risk Sources -- Privacy Metrics -- Sensitivity -- Visibility -- Reachability -- Data Inference -- Suggestion of Countermeasures -- Attribute Visibility in OSN -- Visibility Matrix -- Construction of Visibility Matrix -- An Illustration -- Open Problems -- Harm Trees for OSNs -- Harm Trees -- Construction of Harm trees -- Harm Likelihood -- Harm Expressions -- Harm Database -- Open Problems -- Privacy Risk Analysis in OSNs -- An Overview -- PrivOSN in Details -- Computation of Accuracy -- Evaluation of Harm Likelihoods -- Computation Profile Similarity -- Presentation of Privacy Risk to the User -- Residual Risks -- Open Problems -- Social Benefits -- Social Ties -- Social Capital -- Understanding Social Benefits -- Social Benefit Criteria -- Attributes and Social Benefit Criteria -- Evaluation of Social Benefit -- Open Problems -- Choosing the Right Privacy Settings -- Privacy and Social Benefit Tradeoff in Privacy Management -- An Integer Programming Model -- Balancing Privacy Risks and Social Benefits -- Formulation of the IP Problem -- Objective Function -- Privacy Risk Constraints -- Social Benefit Constraints -- Privacy Setting Constraints -- Open Problems -- Conclusion -- Notations and Their Meanings -- Comparison of Privacy Scoring Mechanisms -- Cases 3 and 4 -- Bibliography -- Authors' Biographies -- Blank Page.

International audience ; Social media such as Facebook provides a new way to connect, interact and learn. Facebook allows users to share photos and express their feelings by using comments. However, Facebook users are vulnerable to attribute inference attacks where an attacker intends to guess private attributes (e.g., gender, age, political view) of target users through their online profiles and/or their vicinity (e.g., what their friends reveal). Given user-generated pictures on Facebook, we explore in this paper how to launch gender inference attacks on their owners from pictures meta-data composed of: (i) alt-texts generated by Facebook to describe the content of pictures, and (ii) comments posted by friends, friends of friends or regular users. We assume these two meta-data are the only available information to the attacker. Evaluation results demonstrate that our attack technique can infer the gender with an accuracy of 84% by leveraging only alt-texts, 96% by using only comments, and 98% by combining alt-texts and comments. We compute a set of sensitive words that enable attackers to perform effective gender inference attacks. We show the adversary prediction accuracy is decreased by hiding these sensitive words. To the best of our knowledge, this is the first inference attack on Facebook that exploits comments and alt-texts solely.

International audience ; Social media such as Facebook provides a new way to connect, interact and learn. Facebook allows users to share photos and express their feelings by using comments. However, Facebook users are vulnerable to attribute inference attacks where an attacker intends to guess private attributes (e.g., gender, age, political view) of target users through their online profiles and/or their vicinity (e.g., what their friends reveal). Given user-generated pictures on Facebook, we explore in this paper how to launch gender inference attacks on their owners from pictures meta-data composed of: (i) alt-texts generated by Facebook to describe the content of pictures, and (ii) comments posted by friends, friends of friends or regular users. We assume these two meta-data are the only available information to the attacker. Evaluation results demonstrate that our attack technique can infer the gender with an accuracy of 84% by leveraging only alt-texts, 96% by using only comments, and 98% by combining alt-texts and comments. We compute a set of sensitive words that enable attackers to perform effective gender inference attacks. We show the adversary prediction accuracy is decreased by hiding these sensitive words. To the best of our knowledge, this is the first inference attack on Facebook that exploits comments and alt-texts solely.

International audience ; To increase awareness about privacy threats, we have designed a tool, SONSAI, for Facebook users to audit their own profiles. SONSAI predicts values of sensitive attributes by machine learning and identifies user public attributes that have guided the learning algorithm towards these sensitive attribute values. Here, we present new aspects of the system such as the automatic combination of link disclosure attacks and attribute prediction. We explain how we defined sensitive subjects from a survey. We also show how the extended tool is fully interfaced with Facebook along different scenarios. In each case a dataset was built from real profiles collected in the user neighbourhood network. The whole analysis process is performed online, mostly automatically and with accuracy of 0.79 in AUC when inferring the political orientation.

International audience ; To increase awareness about privacy threats, we have designed a tool, SONSAI, for Facebook users to audit their own profiles. SONSAI predicts values of sensitive attributes by machine learning and identifies user public attributes that have guided the learning algorithm towards these sensitive attribute values. Here, we present new aspects of the system such as the automatic combination of link disclosure attacks and attribute prediction. We explain how we defined sensitive subjects from a survey. We also show how the extended tool is fully interfaced with Facebook along different scenarios. In each case a dataset was built from real profiles collected in the user neighbourhood network. The whole analysis process is performed online, mostly automatically and with accuracy of 0.79 in AUC when inferring the political orientation.

International audience ; Social media data has been mined over the years to predict individual sensitive attributes such as political and religious beliefs. Indeed, mining such data can improve the user experience with personalization and freemium services. Still, it can also be harmful and discriminative when used to make critical decisions, such as employment. In this work, we investigate social media privacy protection against attribute inference attacks using machine learning explainability and adversarial defense strategies. More precisely, we propose FOX (FOoling with eXplanations), an adversarial attack framework to explain and fool sensitive attribute inference models by generating effective adversarial reactions. We evaluate the performance of FOX with other SOTA baselines in a black-box setting by attacking five gender attribute classifiers trained on Facebook pictures reactions, specifically (i) comments generated by Facebook users excluding the picture owner, and (ii) textual tags (i.e., alttext) generated by Facebook. Our experiments show that FOX successfully fools (about 99.7% and 93.2% of the time) the classifiers, outperforms the SOTA baselines and gives a good transferability of adversarial features.

International audience ; Social media data has been mined over the years to predict individual sensitive attributes such as political and religious beliefs. Indeed, mining such data can improve the user experience with personalization and freemium services. Still, it can also be harmful and discriminative when used to make critical decisions, such as employment. In this work, we investigate social media privacy protection against attribute inference attacks using machine learning explainability and adversarial defense strategies. More precisely, we propose FOX (FOoling with eXplanations), an adversarial attack framework to explain and fool sensitive attribute inference models by generating effective adversarial reactions. We evaluate the performance of FOX with other SOTA baselines in a black-box setting by attacking five gender attribute classifiers trained on Facebook pictures reactions, specifically (i) comments generated by Facebook users excluding the picture owner, and (ii) textual tags (i.e., alttext) generated by Facebook. Our experiments show that FOX successfully fools (about 99.7% and 93.2% of the time) the classifiers, outperforms the SOTA baselines and gives a good transferability of adversarial features.

International audience ; Social media data has been mined over the years to predict individual sensitive attributes such as political and religious beliefs. Indeed, mining such data can improve the user experience with personalization and freemium services. Still, it can also be harmful and discriminative when used to make critical decisions, such as employment. In this work, we investigate social media privacy protection against attribute inference attacks using machine learning explainability and adversarial defense strategies. More precisely, we propose FOX (FOoling with eXplanations), an adversarial attack framework to explain and fool sensitive attribute inference models by generating effective adversarial reactions. We evaluate the performance of FOX with other SOTA baselines in a black-box setting by attacking five gender attribute classifiers trained on Facebook pictures reactions, specifically (i) comments generated by Facebook users excluding the picture owner, and (ii) textual tags (i.e., alttext) generated by Facebook. Our experiments show that FOX successfully fools (about 99.7% and 93.2% of the time) the classifiers, outperforms the SOTA baselines and gives a good transferability of adversarial features.