@article{Wang2026,
  author  = {Wang, Yunlong and Luo, Zhengquan and Zhang, Kunbo and Sun, Zhenan and He, Zhaofeng and Tao, Dacheng and Tan, Tieniu},
  title   = {Synergizing Disentangled Representation and Split Federated Learning to Break Centralized Bottlenecks in Gaze Estimation},
  journal = {International Journal of Computer Vision},
  year    = {2026},
  volume  = {134},
  number  = {11},
  pages   = {492},
  issn    = {1573-1405},
  doi     = {10.1007/s11263-026-03084-8},
  url     = {https://doi.org/10.1007/s11263-026-03084-8},
  abstract = {Mainstream appearance-based gaze estimation methods trained on centralized data are struggling in two urgent aspects. First, most gaze estimation methods require entire eye images or even full faces for model training, which raises serious concerns about the leakage of sensitive personal data. Second, simultaneously achieving generalization to unseen gaze data and personalization for individual subjects remains inherently conflicting to reconcile within a single centralized model. To break these bottlenecks, we propose SplitFedGaze, a distributed gaze estimation framework that synergizes the power of representation disentanglement and split federated learning. Instead of training a global model on centralized data, SplitFedGaze adopts a data-distributed, model-split optimization paradigm. Specifically, gaze-irrelevant components, which primarily involve personal privacy, are removed from the raw gaze data by a carefully designed disentangled information bottleneck module and are retained only on local clients. The desensitized representations, termed ``gaze maps'', encapsulate disentangled gaze-relevant information and are transmitted from local clients to a global server for gaze knowledge aggregation. To resolve the trade-off between generalization and personalization, the server-side gaze knowledge aggregator leverages gaze maps from all clients to enhance the global model's generalizability. Furthermore, each local client is equipped with an exclusive personalized calibration module to mitigate performance degradation caused by physiological differences across subjects. Extensive experiments demonstrate that SplitFedGaze gains three remarkable advantages: (1) It achieves higher estimation accuracy than centralized and federated gaze estimation methods across various training settings. (2) It effectively mitigates the risk of gaze data leakage from reconstruction attacks. (3) It inherently supports personalized applications, overcoming inter-subject variability. Complementary experiments quantify deployment cost, privacy-utility trade-offs, few-shot adaptation, cross-dataset generalization, and update-level privacy mechanisms via differential privacy (DP) and secure aggregation.}
}