Papers by Ishrak Hayet
Invernet: An Inversion Attack Framework to Infer Fine-Tuning Datasets through Word Embeddings (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing word embeddings are data intensive and require large-scale training corpus, sufficient training iterations, and high computational capacity. |
| Approach: | They propose a framework that infers context distributions from a downstream dataset and then uses them to fine-tune the embedding. |
| Outcome: | The proposed framework materializes privacy concern by inferring context distribution in the downstream dataset, which can lead to key information breach. |