AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

Tools for Verifying Neural Models' Training Data

Dami Choi, Yonadav G Shavit, David Duvenaud

NeurIPS 2023 · Adversarial robustness and security › Attacks on text and graphs · approach: Membership inference and inversion · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use memorization test with 1% of data to detect glued runsmemorization plot · GPT-2 gluing attacklong-term memorizationImproved
use memorization test on interpolated checkpointsmemorization plot · GPT-2 interpolation attack, 1% datamemorization diagonalImproved
use memorization test to detect data additionmemorization test · GPT-2 data addition, 1% samplesmemorizationNo clear effect
detect data addition by retraining suspicious segmentssegment retraining · GPT-2 data addition (50% same distribution or 5% Github)weight distance after retrainingImproved
use subtraction upper bound lambda with 1% of datalambda test · GPT-2 data subtraction 0%, 5%, 50%, 95%lambda upper boundImproved
increase model size from Pythia 70M to 1Bmemorization effect · Pythiamemorization clarityImproved
reduce data sampling rate to 1% or 0.1% (ablation)memorization plot · GPT-2 and Pythiamemorization effect visibility: 1% and 0.1%Mixed
later training checkpoints in Pythia 70Mmemorization plot · Pythia 70M with low samplingmemorization effect visibilityHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy