Reproduce “LIA: Privacy-Preserving Data Quality Evaluation in Federated Learning Using a Lazy Influence Approximation” on OpenResearch

About this paper

Ljubomir Rokvic, Panayiotis Danassis, Sai Praneeth Karimireddy, Boi Faltings

A framework called Lazy Influence Approximation (LIA) enables efficient and privacy-preserving evaluation and filtering of low-quality data in Federated Learning environments. It achieves robust data quality assurance by leveraging approximate influence signals and strong differential privacy for both training and validation data, thereby improving model accuracy in diverse settings.

Key insights

• Approximating the *sign* of a data batch's influence, rather than its exact magnitude, is sufficient for effective data quality filtering and drastically reduces computational and communication overhead.
• The combination of DP-SGD for training data privacy and Randomized Response for validation data privacy provides robust, quantifiable differential privacy for the entire quality assessment process.
• LIA is model-agnostic, scalable, and can effectively distinguish between truly harmful corrupted data and valuable data from minority distributions, enhancing FL robustness without discarding rare, important information.

Summary by alphaXiv.