The Internet has become essential to our lives and economies. In particular, it supports a host of personal-data-based online services that we use in almost all our daily activities (for recommendations, social networking, etc.). Thanks to the collection of large amounts of personal data about individuals, such services can provide high value for their users. Despite their success, however, personal-data-based online services increasingly suffer security and privacy problems that strongly affect the trust of users and their willingness to release data, hence holding back the great potential of these services.
Personal-data-based online services rely on learning algorithms to (i) learn from personal data to deliver personalized services and (ii) implement security defenses. In both cases, they use standard learning algorithms developed for data that is independent from the algorithm. Personal data is, however, a very special type of data: it is provided by human agents who can strategically obfuscate it in order to protect their privacy. As a result, standard learning algorithms are unable to combine user privacy and good learning accuracy. Similarly, security data is generated by strategic attackers to evade detection, which renders standard learning algorithms inefficient.
The broad objective of CONNECTED is to design a new generation of secure and private personal-data-based online services, through the invention of learning algorithms and incentive mechanisms that take into account the strategic nature of data provided or generated by humans to achieve higher performance in practice. To achieve this goal, we propose innovative game-theoretic models involving learning that take into account the agents incentives. We solve the games and use the solutions to design game-theoretic statistical learning methods to optimally learn from personal data and secure the system. We implement and test our new algorithms in a real-world service providing ad analytics.