Machine Text Detectors are Membership Inference Attacks
Date:
Abstract
In this talk I discuss my recent joint work with Ryuto Koike on the fundamental link between machine-generated text detection and membership inference. We prove that, for a model that perfectly learns to replicate its training distribution, the optimal metric for both membership inference and machine-generated text detection is the same. For our empirical contribution we demonstrate that many popular methods from both tasks exhibit a high degree of cross-task transfer with rank correlation of 0.66. We also show that the detector Binoculars achieves state-of-the-art performance on membership inference attacks.
Location
This talk was given on Feb 4th, 2025 at the Google Privacy ML Seminar.