Advancements in Differential Privacy for Machine Learning

The field of differential privacy is continuing to evolve, with a focus on developing new methods and techniques to balance privacy and utility in machine learning applications. Recent research has explored the use of correlated noise mechanisms, adaptive gradient clipping, and private evolution algorithms to improve the privacy-utility tradeoff. Notable advancements include the development of algorithms for private relational learning, private synthetic data generation, and private trajectory generation. These innovations have the potential to enable more widespread adoption of differential privacy in real-world applications. Noteworthy papers include: PCEvolve, which proposes a novel algorithm for private contrastive evolution, enabling the generation of high-quality differentially private synthetic images. FERRET, which introduces a fast and effective restricted release mechanism for ethical training, achieving state-of-the-art results in private deep learning.

Sources

PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs

FERRET: Private Deep Learning Faster And Better Than DPSGD

On Differential Privacy for Adaptively Solving Search Problems via Sketching

Correlated Noise Mechanisms for Differentially Private Learning

Private Evolution Converges

Differentially Private Relational Learning with Entity-level Privacy Guarantees

What is the Cost of Differential Privacy for Deep Learning-Based Trajectory Generation?

Built with on top of