From Scarcity to Coverage: Generative Behavioral Modeling for Child Identification on Smartphones

The rising use of mobile devices by minors necessitates robust child identification to mitigate risks like gaming addiction. Existing works mainly focus on continuous identification using children's behavioral data during smartphone usage. However, child behavioral identification is hindered by a critical bottleneck: the scarcity of large-scale, ecologically valid datasets due to high costs and stringent privacy regulations. To address these challenges, this paper proposes a generative model named BehaviorGAN, which synthesizes high-fidelity behavioral data by integrating fine-grained age-group conditioning, downstream task guidance, and physical constraint filtering, for accurate child identification. Experiments on a dataset of 2,255 samples demonstrate that BehaviorGAN serves as a high-utility proxy for real-world behavioral distributions. Moreover, as a data augmentation tool, it improves identification accuracy by 4.5 % and 2.4 % across diverse architectures, providing a scalable and privacy-aware solution for child protection.