Temporal membership inference attacks on continually updated models
Original title: Timing is Everything: Inferring Temporal Membership in Continually Updated Models
TMIA black-box attack infers whether and when samples were used in training for language and image models
Continual model updating creates new privacy risks. Researchers formalize temporal membership inference attack and propose TMIA using bidirectional LSTMs to infer membership at timestamp or interval levels. Experiments on three language and three image classification models show high effectiveness. Defenses evaluated revealing privacy-utility trade-off.
- TMIA significantly outperforms existing baselines
- tested on three language models and three image classification models
- three potential defenses evaluated
- supports timestamp-level and interval-level granularities
From the original article
1 min read at arxiv.orgContinual model updating has become a common practice in the deployment of machine learning systems. However, the dynamic training behaviors give rise to new risks not sufficiently addressed by existing studies. In this paper, we present the first systematic study of temporal privacy risk in continually updated deep learning models.
By Jie Fu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.