Satellite communications (Satcoms) have become an indispensable part of modern society, enabling various applications ranging from global connectivity to critical disaster management. However, reliance on Satcoms also raises concerns about cyber threats, highlighting the need for robust detection mechanisms. This article presents a novel model for threat detection (TD) in Satcoms, leveraging deep-federated learning (DFL). Our DFL-based TD model utilizes variational autoencoders (VAEs) as local models, strategically deployed across satellite ground stations and communication nodes. This architecture is meticulously tailored to the distributed nature of Satcom systems, ensuring robust data privacy at local nodes while enhancing collective TD capabilities. To protect sensitive information during model updates, we integrate differential privacy and secure aggregation techniques, reinforcing the model’s commitment to data confidentiality. Furthermore, we prioritize trustworthiness and accountability in TD processes by emphasizing responsible AI practices, explainability, and ethical compliance. By incorporating SHAP values and local explanation methods, we significantly enhance transparency in decision making, empowering stakeholders to effectively understand and interpret model outputs. In addition, our model addresses bias mitigation and fairness concerns, striving for equitable treatment across diverse data subsets and promoting inclusivity. Real-world case studies demonstrate the model’s effectiveness in detecting anomalies in Satcoms, supporting predictive maintenance, and enhancing network security.

Satellite communications (Satcoms), whether ground-to-space or intersatellite, have, to date, primarily combined the two cutting-edge fields of communication and space, in which cybersecurity and space activities are intricately connected. The susceptibility of Satcoms and other space assets to cyberattacks, as a means of targeting critical infrastructure, is sometimes overlooked. Neither space nor cybersecurity policies, particularly for Satcoms systems, are equipped for the challenges posed by the convergence of space and cyberspace. With the rising number of cyberattacks, including reconnaissance, Denial-of-Service (DoS), and zero-day attacks, on Satcoms systems by extreme nation-state attackers, further defenses must be established. These defenses should enable the discovery of zero-day attacks whilst reducing false positives associated with the detection of advanced persistent threats. In this article, to address these cyber concerns, a comprehensive deep federated learning (DFL)-based threat detection model for proactively recognizing intrusions in Satcoms networks using decentralized on-device data while preserving the privacy of this data is proposed. Our approach leverages a decentralized data-level preprocessing (DLP) mechanism, ensuring that original data remains concealed while providing well-processed, statistically transformed thought-out data for robust threat detection. The proposed model executes federated learning rounds on a novel deep auto-encoder (DAE) architecture, maintaining local data on secure warehouses, and sharing only the learned weights with the central FL server. Also, its ensemble mechanism aggregates the updates from multiple sources to optimize the accuracy of the global learning model. The experimental results demonstrate that the proposed model outperforms the classic/centralised learning (non-FL) versions in terms of protecting the privacy of local data and providing an optimal accuracy rate for attack detection. Furthermore, using differential privacy (DP)-based DLP as a privacy preservation mechanism, the proposed model exhibits high levels of accuracy and better levels of privacy over the training data.

Social media (SM) 3.0 integrates SM platforms, such as Facebook and Twitter, with the Internet of Things (IoT), and has a great potential to change how we interact with mobile devices, online platforms, and the world around us. This integration with end users produces large-scale and heterogeneous data sources that demand machine learning (ML)-based data analytics for decision-making and to provide security against ML and data privacy attacks. The development of privacy-aware ML models within a federated learning (FL) ecosystem can empower an entire network to learn from data in a decentralized manner. In this article, we propose a differentially privacy blockchain-based explainable FL (DP-BFL) framework by harnessing the ever-evolving power of SM 3.0 networks. This framework permits any Internet empowered device to partake and contribute data to a global privacy preserved model. In this framework, participants will upload the differentially private local updates to the miners of blockchain, where the local updates will be evaluated and rewarded. The experimental results obtained from real-world datasets, namely, SM 3.0 and MNIST, demonstrated that the proposed framework could achieve high utility, enhanced privacy, and elevated efficiency. More Specifically, the experimental analysis of our proposed framework reveals the following two key properties. First, our proposed DP-BFL yields noticeable performance improvements in the applied learning models with high privacy and comparable utility levels, in terms of accuracy and f-measure metrics, to a standard FL and centralized learning approaches under the restriction of privacy preservation. Second, given a certain number of the malicious entities, DP-BFL allowed an enhanced recognition of users' preferences in the SM 3.0 dataset and precise prediction of images' class in the MNIST dataset while mitigating the impact of the malicious entities' poisoned updates. Moreover, as the proposed DP-BFL attains DP on the local model's update, it is considered the same as the standard FL-based setting, along with some kinds of privacy preservation on the uploaded model's updates.

Federated Learning (FL), as an emerging form of distributed machine learning (ML), can protect participants’ private data from being substantially disclosed to cyber adversaries. It has potential uses in many large-scale, data-rich environments, such as the Internet of Things (IoT), Industrial IoT, Social Media (SM), and the emerging SM 3.0. However, federated learning is susceptible to some forms of data leakage through model inversion attacks. Such attacks occur through the analysis of participants’ uploaded model updates. Model inversion attacks can reveal private data and potentially undermine some critical reasons for employing federated learning paradigms. This article proposes novel differential privacy (DP)-based deep federated learning framework. We theoretically prove that our framework can fulfill DP’s requirements under distinct privacy levels by appropriately adjusting scaled variances of Gaussian noise. We then develop a Differentially Private Data-Level Perturbation (DP-DLP) mechanism to conceal any single data point’s impact on the training phase. Experiments on real-world datasets, specifically the social media 3.0, Iris, and Human Activity Recognition (HAR) datasets, demonstrate that the proposed mechanism can offer high privacy, enhanced utility, and elevated efficiency. Consequently, it simplifies the development of various DP-based FL models with different tradeoff preferences on data utility and privacy levels.

Social media has become a vital platform in our daily life, where users can interact with their friends and other people throughout the world. The vast data generated by these platforms is unique in its variety and sensitivity, and although it potentially has significant utility, but also the potential for misuse. Although social media providers apply some existing privacy techniques, such as encryption and anonymization, the techniques cannot achieve a solid level of data privacy while maintaining the highest level of data utility. This paper proposes new Privacy-Encoding (PE) models that contain two-levels of data privacy: 1) data perturbation-based encoding techniques, and 2) data normalization-based scaling techniques. The data perturbation-based encoding techniques involve label encoder and one-hot encoder ones, while data normalization-based scaling techniques include min-max and z-score normalization ones. The aim of the two-levels is to transform original data into perturbed data, along with balancing the high level of data utility using machine learning algorithms. To evaluate the data utility, the proposed models are applied on the adult dataset as well as a simulated social media dataset and the accuracy of the results is compared with several machine learning algorithms. The experiment results reveal that the models could achieve high privacy and utility levels in terms of variance, accuracy and f-measure metrics.