New FIBER optimizer enhances differential privacy for AI training

Researchers have introduced FIBER, a novel differentially private optimizer designed to enhance the performance of models trained with temporally filtered gradients. FIBER addresses the issue of miscalibrated bias corrections in adaptive optimizers like AdamW when noise is filtered, by denoising in innovation space and decoupling observation geometry. This new optimizer demonstrates significant improvements in DP training across vision and language benchmarks, outperforming existing methods under strict privacy constraints. AI

影响 Introduces a new method to improve the performance of differentially private model training, potentially enabling more robust privacy guarantees in AI applications.

排序理由 This is a research paper detailing a new optimization method for differentially private training. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• AdamW
• arXiv
• FIBER

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →