Robust RSSI indoor localization through federated adversarial training

Wi-Fi RSSI-based indoor localization offers a practical foundation for context-aware services but is challenged by data privacy concerns and susceptibility to adversarial perturbations. This paper presents FedAdvLoc as an application-oriented federated training design that incorporates client-side AdvGAN-RSSI-generated adversarial augmentation to enhance robustness while keeping raw RSSI data local during Stage 2 federated deployment/adaptation. During federated deployment/adaptation, each client uses a frozen building-specific AdvGAN-RSSI generator as a structured perturbation source, applies one current-model gradient-sign refinement step, and optimizes a scheduled clean/adversarial mixture using FedProx. Evaluated on the SODIndoorLoc dataset across three heterogeneous buildings under IID and non-IID partitions, FedAdvLoc achieves unperturbed-input localization accuracy of 2.57–4.43 m RMSE and 92.16–95.71% success rate, closely matching robust centralized baselines. Under strong adversarial conditions, IID and non-IID models maintain \\(\\Delta \\) RMSE below 0.6 m and 1.6 m, respectively, reducing RMSE error growth by up to 96.9% compared with the non-robust centralized baseline while retaining approximately 79–95% success rates across FGSM, PGD-RS, and combined AdvGAN-RSSI+FGSM attacks. These results indicate that client-side adversarial augmentation can improve perturbation tolerance within a raw-data-local Stage 2 federated setting, supporting practical and robustness-aware deployment of RSSI-based indoor localization in multi-building, multi-device IoT environments.