Related Experiment Video
Updated: Jan 10, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
994
Adaptive differential privacy mechanism for enhanced deep learning model utility and privacy
Zhang Xiangfei1, Zhang Qingchen2
1School of Cyberspace Security, Hainan University, Haikou, 570228, China.
Summary
This study introduces an adaptive differential privacy (DP) method for deep learning, improving model performance and privacy protection by dynamically adjusting sensitivity and budget allocation during training.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Data Privacy
Background:
- Deep learning models risk training data privacy leakage.
- Differential privacy (DP) with SGD is used for privacy guarantees.
- Current DP methods use constant gradient clipping, impacting performance and requiring tuning.
Purpose of the Study:
- To propose a novel deep learning strategy with an adaptive DP mechanism.
- To achieve adaptive sensitivity and adaptive privacy budget allocation.
- To improve the trade-off between model utility and privacy protection.
Main Methods:
- Clustering average batch gradients to adaptively allocate privacy budgets.
- Coupling gradient clipping parameters with the learning rate for adaptive clipping.
- Rigorous theoretical privacy analysis to ensure DP compliance.
Main Results:
- The proposed adaptive DP method strictly satisfies DP guarantees.
- Outperforms state-of-the-art approaches in visual tasks.
- Avoids manual hyperparameter tuning for gradient clipping.
Conclusions:
- Adaptive DP offers enhanced privacy and utility in deep learning.
- The novel strategy effectively manages privacy-budget trade-offs.
- Demonstrates superior performance in mainstream visual tasks.
