AI Paper Atlas

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

Fooling Explanations in Text Classifiers

Adam Ivankay, Ivan Girardi, Chiara Marchiori, Pascal Frossard

ICLR 2022 · Adversarial robustness and security › Attacks on text and graphs · approach: Defenses against universal perturbations · 5 citations (OpenAlex)

Read the paper

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 TEF's attribution-guided word ranking and best-synonym selection instead of RandomAttack's random choicesTextExplanationFooler (TEF) · Five text-classification datasets, tested across the reported models and explainersattribution correlationImproved
Randomize the word-importance ranking while retaining best-synonym selection (ablation)TEF · Text-classification datasets and models in the ablation studyPCCHurt
Randomize final synonym selection while retaining word-importance ranking (ablation)TEF · Text-classification datasets and models in the ablation studyPCCImproved
Transfer TEF perturbations from another model or explanation method rather than optimize them on the targetTEF · Text-classification models and explanation methodsPCC: approx. 0.4Mixed
Apply precomputed semi-universal replacement policies without querying the target during attack timeTEF · Text-classification models and explanation methodsPCCImproved
Increase the maximum perturbed-token ratioTEF · Text-classification datasets and modelsattribution correlationImproved

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