CA-ADP: Adaptive Differential Privacy for Sensor-Based Fall Detection
Researchers propose a Class-Aware Adaptive Differential Privacy (CA-ADP) framework for sensor-based fall detection using deep learning. The method integrates with a hybrid 3D CNN-BiLSTM architecture to dynamically adjust noise based on class composition in each mini-batch, improving privacy-utility trade-off compared to uniform noise approaches. The work addresses privacy concerns in healthcare sensor data and provides a formal (ε,δ)-differential privacy guarantee. The paper is available on arXiv (2605.01679).
Key facts
- Fall detection is critical for elderly healthcare.
- Sensor-based activity data raises privacy concerns.
- Existing privacy approaches add uniform noise, degrading performance.
- CA-ADP adjusts noise magnitude based on class composition.
- Framework uses hybrid 3D CNN-BiLSTM architecture.
- Provides (ε,δ)-Differential Privacy guarantee.
- Paper published on arXiv with ID 2605.01679.
- Aims to balance privacy and utility.
Entities
Institutions
- arXiv