Combining Homomorphic Encryption and Differential Privacy in Federated Learning for Model Inspection and Availability
A federated-learning design combines homomorphic encryption and differential privacy, beating DP-only utility and privacy on FEMNIST.
The paper proposes federated learning that trains under homomorphic encryption while using differential privacy for model inspection and release. Privacy is estimated with a Markov chain Monte Carlo Bayesian method. On FEMNIST, the approach reached a test loss of 1.09 versus 2.37 for differential-privacy-only training, with an estimated posterior mean epsilon of 4.32 versus 7.26. Intermittent monitoring preserved the encrypted training trajectory while matching or improving estimated privacy under the tested setting.
- Training uses homomorphic encryption; differential privacy covers inspection and release.
- A Bayesian MCMC method estimates privacy instead of DP-only training.
- On FEMNIST, test loss was 1.09 versus 2.37 for the DP-only baseline.
- Estimated epsilon posterior mean was 4.32 versus 7.26 for that baseline.
From the original article
1 min read at arxiv.orgThe increasing prevalence of decentralized data has led to a growing interest in federated learning, which enables collaborative model training without clients sharing their sensitive local data. However, FL alone does not sufficiently protect sensitive training data and is generally coupled with privacy-preserving techniques, such as differential privacy and homomorphic encryption.
By Ceren Yıldırım · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.