Papers by Fatemehsadat Mireshghallah
Style Pooling: Automatic Text Style Obfuscation for Improved Classification Fairness (2021.emnlp-main)
Copied to clipboard
| Challenge: | Text style can reveal sensitive attributes of the author (e.g. age and race) to the reader, which can lead to privacy violations and bias in both human and algorithmic decisions based on text. |
| Approach: | They propose a framework that obfuscates stylistic features of human-generated text through style transfer by automatically re-writing the text itself. |
| Outcome: | The proposed framework obfuscates stylistic features of human-generated text through style transfer, by automatically re-writing the text itself. |
Quantifying Privacy Risks of Masked Language Models Using Membership Inference Attacks (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prior attempts at measuring leakage of MLMs via membership inference attacks have been inconclusive, implying potential robustness of Mlms to privacy attacks. |
| Approach: | They propose a stronger membership inference attack based on likelihood ratio hypothesis testing that involves an additional MLM to more accurately quantify the privacy risks of memorization in MLMs. |
| Outcome: | The proposed attack improves the AUC of prior membership inference attacks from 0.66 to an alarmingly high 0.90 level on models trained on medical notes. |
Privacy-Preserving Natural Language Processing (2023.eacl-tutorials)
Copied to clipboard
| Challenge: | This tutorial will help the NLP community to get familiar with current research in privacy-preserving methods. |
| Approach: | This tutorial will help the NLP community to get familiar with current research in privacy-preserving methods. |
| Outcome: | The tutorial will cover membership inference, differential privacy, homomorphic encryption, or federated learning, all with typical use-cases and potential pitfalls. |
An Empirical Analysis of Memorization in Fine-tuned Autoregressive Language Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Large language models are shown to present privacy risks through memorization of training data, but little attention has been given to the fine-tuning phase. |
| Approach: | They empirically study memorization of fine-tuning methods using membership inference and extraction attacks and show that fine-timing the head of the model has the highest susceptibility to attacks. |
| Outcome: | The proposed methods have the highest memorization risk, whereas the smaller adapters are less vulnerable to known extraction attacks. |
UserIdentifier: Implicit User Representations for Simple and Effective Personalized Sentiment Analysis (2022.naacl-main)
Copied to clipboard
Fatemehsadat Mireshghallah, Vaishnavi Shrivastava, Milad Shokouhi, Taylor Berg-Kirkpatrick, Robert Sim, Dimitrios Dimitriadis
| Challenge: | Currently, global models are not able to produce personalized responses for individual users, based on their data. |
| Approach: | They propose a scheme for training a single shared model for all users by prepending a fixed, user-specific non-trainable string to each user’s input text. |
| Outcome: | The proposed method outperforms the state-of-the-art model on a suite of sentiment analysis datasets by up to 13 points. |
Membership Inference Attacks against Language Models via Neighbourhood Comparison (2023.findings-acl)
Copied to clipboard
Justus Mattern, Fatemehsadat Mireshghallah, Zhijing Jin, Bernhard Schoelkopf, Mrinmaya Sachan, Taylor Berg-Kirkpatrick
| Challenge: | Existing membership inference attacks aim to predict whether a data sample was present in training data of a machine learning model. |
| Approach: | They propose to compare model scores to neighbour texts to eliminate access to training data by comparing model scores with a given sample. |
| Outcome: | The proposed attacks outperform reference-based attacks with perfect knowledge of the training data distribution and outperformed reference-free attacks with imperfect knowledge. |
Mix and Match: Learning-free Controllable Text Generationusing Energy Language Models (2022.acl-long)
Copied to clipboard
| Challenge: | Recent work on controlled text generation has required attribute-based fine-tuning of the base language model or restricted the parameterization of the attribute discriminator. |
| Approach: | They propose a global score-based alternative for controllable text generation that combines arbitrary pre-trained black-box models for achieving desired attributes in the generated text. |
| Outcome: | The proposed method outperforms methods that require extra training or fine-tuning . the proposed method is based on a model with energy values of a linear combination of scores from black-box models . |
Privacy Regularization: Joint Privacy-Utility Optimization in LanguageModels (2021.naacl-main)
Copied to clipboard
Fatemehsadat Mireshghallah, Huseyin Inan, Marcello Hasegawa, Victor Rühle, Taylor Berg-Kirkpatrick, Robert Sim
| Challenge: | Neural language models have a high capacity for memorization of training samples . however, this can cause privacy degradation and disparate impact on subgroups of users . |
| Approach: | They propose two privacy-preserving regularization methods for training language models that enable joint optimization of utility and privacy. |
| Outcome: | The proposed methods have favorable utility-privacy trade-off, faster training and uniform treatment of under-represented subgroups. |
Privacy-Preserving Domain Adaptation of Semantic Parsers (2023.acl-long)
Copied to clipboard
| Challenge: | Task-oriented dialogue systems often assist users with personal or confidential matters . a lack of privacy controls prevents developers from observing actual usage . authors propose a method to generate realistic user utterances synthetically without compromising privacy . |
| Approach: | They propose a method which generates latent semantic parses and generates utterances based on the parses. |
| Outcome: | The proposed method improves MAUVE by 2.5X and parse tree function-type overlap by 1.3X . it also shows gains of 8.5% points on its accuracy with the new feature . |