Minghui Xu

2 previous affiliations
2Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Minghui Xu focuses on privacy preservation in large language models and the structural processing of spoken language. Their work investigates methods for detecting and editing memorized private information within neural networks, alongside the development of resources for spoken-to-written style conversion. Recent efforts address the mitigation of data leakage risks without requiring full model retraining.

Research Trajectory

In 2023, Minghui Xu's research primarily addressed vulnerabilities in large language models and the challenges of processing transcribed speech. A significant portion of this work falls within the research direction of Selective Knowledge Erasure in Pretrained Models. As models scale, they increasingly memorize sensitive training data, creating risks of unintended data leakage that are often only discovered after training is complete. Participating as a co-author, Xu contributed to DEPN: Detecting and Editing Privacy Neurons in Pretrained Language Models, which introduced a framework to mitigate these risks without the prohibitive computational cost of retraining from scratch. The study proposed a "privacy neuron detector" to locate specific network components associated with private information, subsequently editing them by setting their activations to zero. This work intersects with the broader area of Data Reconstruction and Membership Privacy in Deep Learning, emphasizing targeted interventions to cure models rather than discarding them.

Alongside privacy research, Xu has contributed to natural language processing resources aimed at improving downstream task performance on transcribed text. As a co-author, Xu helped develop CS2W: A Chinese Spoken-to-Written Style Conversion Dataset with Multiple Conversion Types. Because automatic speech recognition outputs often contain disfluencies, colloquialisms, and grammatical errors, converting spoken language into a formal written style is necessary for robust text processing. The CS2W dataset provides over 7,000 annotated spoken sentences extracted from conversational texts to facilitate research into these multiple conversion problems.

Current Directions

The most recent papers in the evidence, published in 2023, indicate a focus on model editing for privacy preservation and the refinement of noisy linguistic data. In the domain of language model security, this work addresses the tension between model capabilities and data memorization, aligning with themes such as The Privacy-Utility Trade-off in Differentially Private Learning. Concurrently, efforts in dataset creation highlight an interest in bridging the gap between raw spoken transcriptions and structured written formats to improve the reliability of downstream natural language processing tasks.

Profile updated

See something wrong? Report a mistake