Location-Based Recommendation Systems (LBRS) use device location data to suggest nearby hotels, restaurants, and points of interest. Since directly collecting location data from users can raise privacy concerns, there is growing interest in building recommendation systems based on Federated Learning (FL). Under FL, parameters of recommendation model learned on each user’s device are collected on a single server to build aggregated model. While FL does not raise privacy concerns about data collection since it does not collect user data directly, it may construct unfair models that repeatedly recommend specific locations. Although there are training methods to achieve fair recommendations that prevent such bias, they require more training epochs than usual. In FL, a malicious server can infer the original location data by continuously tracking a specific user’s parameter updates, and the inference accuracy increases proportionally with the number of training epochs. This means that achieving fair location recommendations in FL puts the original data at risk. In this paper, we design a novel parameter aggregation method to build fair and secure FL recommendation models. In the proposed aggregation method, users exchange parameters with each other before model aggregation to prevent malicious servers from inferring the original data. Even if a server (adversary) continuously tracks a specific user’s device, it cannot get parameters from the same user, thus preventing inference of the original location data. An experiment result demonstrated that the proposed method can reduce training time while maintaining the same accuracy as homomorphic encryption approach. © 2013 IEEE.

Spatio-temporal data possess intrinsic values, reflecting the spatial and temporal features of people's behaviors. Due to the sensitive nature of this data (e.g., workplace, residence, school locations), privacy protection is essential when collecting spatio-temporal data. Local Differential Privacy (LDP) protocol has gained attention as a method for protecting privacy on data-collecting devices. LDP protocol can make each data indistinguishable but inevitably destroys spatial/temporal characteristics as well. In this paper, we propose a novel method enabling LDP protocol to preserve spatial/temporal trends on privacy protection. If we collect data from users with similar behavior, it is difficult to uniquely identify users from the beginning. In short, processing privacy protection for each user with similar behavior allow us to minimize the removal of intrinsic values by LDP protocol. Our method, termed Dynamic Differentially-Private Spatial Decomposition (D-2-PSD), dynamically adjusts and controls the strength of privacy protection (privacy budget) for each group of users exhibiting similar spatial and temporal trends. This allows users to be indistinguishable from each other within a group while preserving spatial and temporal trends across groups. All groups will have a different privacy budget, but the sum of the entire group keeps a constant privacy budget. Even if group with different protection strengths are mixed, privacy is protected for the sum of the group, and our proposed method can always guarantee a constant protection strength. Experimental results demonstrate that our method retains the intrinsic spatial and temporal trends in spatio-temporal data while maintaining robust privacy protection across the entire dataset, thanks to the D-2-PSD approach. Specifically, in the most similar groups, D-2-PSD reduced the MAE by up to 75% compared to standard LDP, while maintaining an equivalent strength of privacy protection.

The digitization of financial transactions has led to a rise in credit card fraud, necessitating robust measures to secure digital financial systems from fraudsters. Nevertheless, traditional centralized approaches for detecting such frauds, despite their effectiveness, often do not maintain the confidentiality of financial data. Consequently, Federated Learning (FL) has emerged as a promising solution, enabling the secure and private training of models across organizations. However, the practical implementation of FL is challenged by data heterogeneity among institutions, complicating model convergence. To address this issue, we propose FedFusion, which leverages the fusion of local and global models to harness the strengths of both, ensuring convergence even with heterogeneous data with total feature discrepancy. Our approach involves three distinct datasets with completely different feature sets assigned to separate federated clients. Prior to FL training, datasets are preprocessed to select significant features across three deep learning models. The Multilayer Perceptron (MLP), identified as the best-performing model, undergoes personalized training for each dataset. These trained MLP models serve as local models, while the main MLP architecture acts as the global model. FedFusion then adaptively trains all clients, optimizing fusion proportions. Experimental results demonstrate the approach's superiority, achieving detection rates of 99.74%, 99.70%, and 96.61% for clients 1, 2, and 3, respectively. This highlights the effectiveness of FedFusion in addressing data heterogeneity challenges, thereby paving the way for more secure and efficient fraud detection systems in digital finance.