The field of machine learning is moving towards a greater emphasis on privacy preservation, with several recent papers exploring innovative methods for protecting sensitive data. One key direction is the development of differentially private fine-tuning methods for large language models, which can help prevent the exposure of sensitive training data. Another area of focus is the creation of robust out-of-distribution detection systems, which can help identify and mitigate potential privacy risks. Additionally, researchers are exploring the use of graph foundation models and large language models to enable zero-shot graph out-of-distribution detection and graph synthetic out-of-distribution exposure. Noteworthy papers include NoEsis, which proposes a framework for differentially private knowledge transfer in modular LLM adaptation, and ReCIT, which presents a novel privacy attack for reconstructing full private data from gradients in parameter-efficient fine-tuning of LLMs. GLIP-OOD is also notable for its framework for zero-shot graph OOD detection using a graph foundation model and LLMs.
Advances in Privacy-Preserving Machine Learning
Sources
NoEsis: Differentially Private Knowledge Transfer in Modular LLM Adaptation
Revisiting Data Auditing in Large Vision-Language Models
Co-Training with Active Contrastive Learning and Meta-Pseudo-Labeling on 2D Projections for Deep Semi-Supervised Learning
Active Few-Shot Learning for Vertex Classification Starting from an Unlabeled Dataset
ReCIT: Reconstructing Full Private Data from Gradient in Parameter-Efficient Fine-Tuning of Large Language Models
Can Differentially Private Fine-tuning LLMs Protect Against Privacy Attacks?
GLIP-OOD: Zero-Shot Graph OOD Detection with Foundation Model
Graph Synthetic Out-of-Distribution Exposure with Large Language Models
Beyond Public Access in LLM Pre-Training Data