CrossDiscipline
Federated Learning + Privacy Meets Epistemology
Towards Explainable Federated Learning: Understanding the Impact of Differential Privacy
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceWhat the paper says
DP harms explainability but the tradeoff is acceptable for the privacy gains.
The Critique
They didn't cite the philosophy of science literature on privacy vs. knowledge trade-offs, particularly debates around informed consent in research and the 'open data' movement. Nor did they reference the work on 'epistemic injustice'—how privacy protections can systematically disadvantage certain groups if the noise patterns interact with demographic distributions.
Why It Matters
If differential privacy fundamentally conflicts with explainability, this creates a dilemma for regulated applications requiring both. Understanding the mechanism of this conflict could guide development of privacy-preserving explainable AI methods.
What They Missed
They don't quantify how DP harms explainability or analyze the mechanism. Is it because DP adds noise to feature importance scores, or because privatized trees select different features?