Original title: Pushing the (Decision) Boundaries: Dynamically Calibrating Differentially Private Noise to Explainability in Federated Learning
Michael Khavkin, Kichang Lee, Jaeho Jin, JeongGil Ko +1 · September 2026
Federated learningA way for multiple institutions to train AI models together without sharing their raw data. lets different devices train an AI model together without sharing their private data. To keep this data safe, engineers often add mathematical noise, known as differential privacyA technique that adds carefully calibrated noise to data or models to protect individual privacy.. However, this noise damages the model's accuracy and ruins its explanations, which shows which parts of an image influenced its choices. Older methods could not adjust this noise to improve explanations during training. To solve this, researchers designed a training method called XCal-FL for image classification. It checks how model confidence shifts, how close samples are to decision boundaries, and how well visual attention stays focused. Using these three signals, it dynamically tunes the added noise while maintaining privacy guarantees. The team tested it on three medical imaging datasets. XCal-FL improved predictive performance by over 10% and boosted explanation fidelity by up to 5x compared to static-noise setups, while spending the privacy budget more efficiently.
Imagine you're compiling a recipe book with friends, but you can't reveal your secret ingredients. Each person writes down steps with random changes to hide the secrets, but this can make the recipes hard to follow. Here, the system fine-tunes how much random change is added based on whether the recipe still tastes good and is easy to explain, balancing secrecy and usefulness.
In healthcare, doctors rely on AI explanations to trust and use models for diagnosis. If privacy protections distort these explanations, it could lead to wrong decisions. This method makes AI both private and interpretable, crucial for sensitive applications.
The abstract doesn't explicitly state limitations, but the work is focused on image classification in cross-silo federated learning for medical imaging. It may not generalize to other domains or types of machine learning tasks.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Pushing the (Decision) Boundaries: Dynamically Calibrating Differentially Private Noise to Explainability in Federated Learning — Michael Khavkin, Kichang Lee, Jaeho Jin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.