In this study we shed the light on the danger of privacy leakage on social network. We investigate privacy breaches, design attacks, show their feasibility and study their accuracies. This approach allows us to track the origin of threats and is a first step toward designing effective countermeasures. We have first introduced a subject sensitivity measure through a questionnaire survey. Then, we have designed on-line friendship and group membership link disclosure (with certainty) attacks on the largest social network ¿Facebook¿. These attacks successfully uncover the local network of a target using only legitimate queries. We have also designed sampling techniques to rapidly collect useful data around a target. The collected data are represented by social-attribute networks and used to perform attribute inference (with uncertainty) attacks. To increase the accuracy of attacks, we have designed cleansing algorithms. These algorithms quantify the correlation between subjects, select the most relevant ones and combat data sparsity. Finally, we have used a shallow neural network to classify the data and infer the secret values of a sensitive attribute of a given target.