ABSTRACT The proposal of PFL (personalized federated learning) for STINs (satellite‐terrestrial integrated networks) is a recent development in the field, with the aim of enhancing the accuracy of models through the utilization of personalized models for each client. These models are derived from diverse devices within the STINs, thereby ensuring a comprehensive and representative dataset. In the context of prevailing PFL schemes, clients receive local models from a server and subsequently aggregate these local models using identical weights. Nevertheless, this equal weighting approach engenders a lower accuracy of the aggregated local models. Conversely, the lower accuracy rate has been observed to trigger DoS (denial‐of‐service) attacks. In this paper, we propose PFL‐Sec, a PFL framework for STINs security, with the aim of optimizing the distribution of weights and improving the model's accuracy against DoS attacks. Specifically, an optimization method of weights for local models based on gradient descent is proposed, with the aim of strengthening the weights of models with high contribution to the model in personalized aggregation. The experimental results indicate that PFL‐Sec outperforms the other three baselines and improves the accuracy by 2.61% under the same settings. Furthermore, PFL‐Sec demonstrates efficacy in the realm of DoS attacks. The experimental results indicate that the success rate of DoS in PFL‐Sec is only 50% when the accuracy threshold for local models is set at 70%. It is evident that clients will only engage in the aggregation process if the accuracy of their local model exceeds 70%.
Wang et al. (Sun,) studied this question.