Paper proposes unified framework for efficient model unlearning in vision and audio

Researchers have introduced Graph-Propagated Projection Unlearning (GPPU), a novel method designed to selectively remove learned information from deep neural networks. This technique is applicable to both vision and audio models, utilizing graph-based propagation to isolate and remove specific class data. Evaluations on multiple datasets and architectures indicate that GPPU offers significant speedups compared to existing methods while maintaining the model's performance on other classes. AI

影响 Provides a more efficient and privacy-preserving method for selectively removing data from AI models.

排序理由 This is a research paper detailing a new unlearning method for AI models.

在 arXiv cs.CV 阅读 →

• Audio Transformer
• CNN
• GPPU
• Graph Propagated Projection Unlearning
• Shreyansh Pathak
• Vision Transformer

• 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 →