Class-Aware Adaptive Differential Privacy in Deep Learning for Sensor-Based Fall Detection

Technical Summary: Class-Aware Adaptive Differential Privacy in Deep Learning for Sensor-Based Fall Detection

Problem Statement
Fall detection is a critical healthcare application for the elderly, where timely intervention can prevent severe injuries. While sensor-based systems using accelerometers and gyroscopes offer advantages over vision-based methods, the data they collect is highly sensitive, containing behavioral and health-related information. Deep learning models trained on this data are vulnerable to privacy attacks, such as membership inference and model inversion.

Existing privacy-preserving approaches, primarily Differentially Private Stochastic Gradient Descent (DP-SGD), address this by injecting uniform noise across all training samples. However, fall detection datasets are inherently imbalanced, with fall events (minority class) being significantly fewer than Activities of Daily Living (ADLs). The authors argue that applying uniform noise disproportionately degrades the learning of minority classes, leading to a significant drop in prediction performance and recall for fall events. There is a lack of frameworks that jointly address spatiotemporal feature learning and privacy preservation while accounting for class imbalance.

Methodology
The paper proposes a Class-Aware Adaptive Differential Privacy (CA-ADP) framework integrated with a hybrid 3D Convolutional Neural Network and Bidirectional Long Short-Term Memory (3D CNN-BiLSTM) architecture.

1. Model Architecture:

   • 3D CNN: Extracts spatiotemporal features from raw sensor signals (accelerometer and gyroscope data) segmented into fixed-length windows.
   • BiLSTM: Processes the extracted features in both forward and backward directions to capture long-term temporal dependencies and contextual information surrounding potential fall events.
   • Output: A fully connected layer with a sigmoid activation for binary classification (Fall vs. Non-Fall).

2. Class-Aware Adaptive Differential Privacy (CA-ADP):

   • Unlike standard DP-SGD which applies a fixed noise multiplier, CA-ADP dynamically adjusts the noise magnitude based on the class composition of each mini-batch.
   • Mechanism: The noise standard deviation ($\sigma_b$) for a mini-batch is calculated as $\sigma_b = \sigma_{base}(1 - \alpha \bar{y}_b)$, where $\bar{y}_b$ is the mean label of the batch (proportion of fall events) and $\alpha$ is an adaptive factor.
   • Effect: Batches dominated by fall events (minority class) receive lower noise injection to preserve the gradient signal, while batches dominated by ADLs (majority class) receive higher noise.
   • Privacy Guarantee: The system maintains a formal $(\epsilon, \delta)$-Differential Privacy guarantee. The privacy budget is calculated using the Rényi Differential Privacy (RDP) accountant, based on the maximum noise value ($\sigma_{base}$) applied across all batches to ensure a conservative upper bound.

3. Experimental Setup:

   • Datasets: The framework was evaluated on three public benchmark datasets: SisFall, UP-Fall, and MobiAct.
   • Preprocessing: A standardized pipeline involving noise filtering (Butterworth low-pass filter for SisFall), normalization (StandardScaler), and sliding window segmentation with 50% overlap.
   • Baselines: Comparisons were made against standalone 3D CNN, standalone BiLSTM, and a hybrid 3D CNN-BiLSTM using conventional uniform DP-SGD (CONV-DP).

Key Results
The proposed CA-ADP framework demonstrated superior performance compared to conventional privacy-preserving models across all three datasets:

• Performance Gains: In terms of F-score, CA-ADP achieved improvements of 3.3% (SisFall), 8.5% (UP-Fall), and 7.5% (MobiAct) over the conventional DP-SGD baseline.
• Recall Improvement: The most significant gains were observed in Recall, with CA-ADP improving by approximately 0.088 on average compared to CONV-DP. This is critical for fall detection, where missing a fall (False Negative) is more detrimental than a false alarm.
• Statistical Significance: Wilcoxon signed-rank tests confirmed that the performance improvements of CA-ADP over CONV-DP are statistically significant ($p < 0.0167$ after Bonferroni correction) across all metrics and datasets.
• Privacy-Utility Trade-off: The analysis showed that CA-ADP achieves a better balance between privacy and utility. On the SisFall and UP-Fall datasets, CA-ADP achieved lower privacy budgets ($\epsilon$) than CONV-DP (indicating stronger privacy protection for the same utility), while on MobiAct, the privacy budget was comparable.
• Comparison with Prior Work: The proposed model achieved competitive or superior accuracy and F-scores compared to previous studies on these datasets, with the added advantage of providing formal privacy guarantees, which prior works largely lacked.

Significance and Claims
The authors claim that this study is the first to apply a Class-Aware Adaptive Differential Privacy mechanism specifically to the fall detection problem.

• Addressing Imbalance: The primary contribution is demonstrating that adaptive noise scheduling based on mini-batch class composition can mitigate the performance degradation typically associated with differential privacy in imbalanced healthcare datasets.
• Formal Guarantees: Unlike many existing fall detection studies that focus solely on accuracy, this work provides rigorous $(\epsilon, \delta)$-DP guarantees, addressing the ethical and legal necessity of protecting sensitive user data in healthcare AI.
• Practical Utility: The results suggest that CA-ADP not only preserves privacy but can act as an implicit regularizer, potentially improving generalization and recall for rare events (falls) better than non-private baselines in certain scenarios.

Limitations and Future Directions
The authors modestly acknowledge that the achieved privacy budgets ($\epsilon > 10$) are not as stringent as those advocated in some differential privacy literature, a consequence of the small batch size and noise multipliers used. They also note that the datasets were collected under controlled laboratory conditions. Future work aims to reduce the privacy budget to the $\epsilon = 3-5$ range, explore zero-concentrated differential privacy, extend the framework to multi-class fall type recognition, and integrate the mechanism into a federated learning setting.