Intelligent Privacy-preserving Approaches for safeguarding Internet of Things-integrated Social Media Networks

Abstract

Web 3.0 represents the third generation of web technologies for incorporating decentralisation and agility in web applications. As it integrates Social Media (SM) in the Internet of Things (IoT), the endless synergies established promise consumers greater connectivity and interaction as well as more seamless movement between physical spaces. Despite these advantages, this integration also increases issues of cyber vulnerabilities and cyberattacks that cause financial, political and social damage in such data-rich environments. Although machine learning underpins this transition, the data used may contain sensitive information that could be compromised by privacy and security breaches. Therefore, techniques that maintain the utility of such valuable data while protecting individuals' privacy are increasingly being required, with federated learning-based privacy preservation ones the current standard. However, as federated learning approaches still require a central authority, they enable significant cyberattacks.

In this thesis, a major contribution is the safeguarding of heterogeneous data, such as that in IoT-integrated SM networks (i.e., SM 3.0). It does so by introducing several new contributions, a novel privacy preservation IoT-integrated SM framework and extended and improved federated learning-based ones that intensively leverage the power of the privacy definition involved in differential privacy and the trustworthiness of blockchain modules to protect against privacy attacks. Apart from improving privacy preservation in conventional federated learning-based frameworks in future social platforms, in this work, accountability and trustworthiness are investigated, with blockchain-based systems integrated into the developed frameworks. Simultaneously, by applying privacy preservation techniques, a preserved and more statistical data version of IoT-integrated SM data is offered to machine learning-based applications to maintain high utility levels.

The experimental results reveal the following two key properties; the proposed frameworks yield noticeable performance improvements with high privacy and comparable utility levels while also allowing enhanced recognition of user preferences in the SM 3.0 dataset developed as part of this work to address the unavailability of data sources as well as highly classifying a user's actions based on observations of the surrounding environment in the IoT datasets. Furthermore, since the proposed frameworks attain differential privacy on the training data, they are considered the same as standard federated learning but with a greater level of privacy preservation. In future, these frameworks will be extended by examining their integration with other distributed learning ones.