CHAMPEIL, M. Nathan (2024) Attaques par Inférence d’Appartenance contre les Grands Modèles de Langage- Membership Inference Attacks against Large Language Models PRE - Research Project, ENSTA.
![]()
| PDF 937Kb |
Abstract
The recent development of large language models raises legal concerns regarding the respect of the copyright. The membership inference attacks might be a means to adjudicate the ongoing lawsuits. We study the dataset used in an article that seemed to obtain promising results, and we show that it contains statistical biases, that these biases skew the mesurements, and that it is hard to rid the dataset of these biases a posteriori.
Item Type: | Thesis (PRE - Research Project) |
---|---|
Uncontrolled Keywords: | Large Language Model, Membership Inference Attack, Dataset, Bias |
Subjects: | Information and Communication Sciences and Technologies |
ID Code: | 10266 |
Deposited By: | M Nathan CHAMPEIL |
Deposited On: | 09 sept. 2024 14:22 |
Dernière modification: | 09 sept. 2024 14:22 |
Repository Staff Only: item control page