M. Ali Fauzi, Bian Yang, Yati Sri Hayati, Budi Darma Setiawan, Irawati Nurmala Sari, Eko Sakti Pramukantoro
Falls are a significant health risk among elderly populations, necessitating effective detection systems. Vision-based methods, particularly pose estimation combined with Long Short-Term Memory (LSTM) networks, have shown promising accuracy but raise privacy concerns due to the use of sensitive visual data. This paper proposes a privacy-aware fall detection system that leverages pose-based skeletal data and integrates Differential Privacy (DP) into LSTM training, providing formal guarantees of data privacy. We evaluate our approach using a publicly available dataset, comparing standard LSTM models against those trained with DP at various privacy levels (privacy budgets, ε). Our results indicate that incorporating DP introduces a measurable performance trade-off: higher privacy protection (smaller ε) leads to reductions in accuracy, precision, recall, and F1-score, particularly affecting the detection of minority-class events (falls). Nevertheless, moderate privacy settings (ε ≈ 2 to 5) still achieve acceptable performance (approximately 85-87% accuracy), balancing privacy and practical utility effectively. This work underscores the importance of carefully selecting privacy budgets to maintain model efficacy while ensuring rigorous data protection, offering a robust approach suitable for sensitive environments such as elderly care facilities. © 2025 IEEE.
Faculty of Computer Science, Universitas Brawijaya, Malang, Indonesia; Department of Information Security and Communication Technology, Norwegian University of Science and Technology (NTNU), Gjøvik, Norway
