AI Paper Atlas

Research map / Federated learning and privacy / Privacy attacks and secure learning

The Hidden Risk: Membership Inference Attacks on Multimodal Federated Learning via Modality Imbalance

chang Ma, Jun Li, Kang Wei, Yipeng Zhou, Ming Ding, Yiyang Ni

ICML 2026 · Federated learning and privacy › Privacy attacks and secure learning · approach: Defending federated learning against poisoning

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
attack only the dominant (audio) modality instead of the whole multimodal model (ablation)Cosine-LIRA · CREMA-DAUC: 0.7500 -> 0.7551No clear effect
attack only the non-dominant (visual) modality (ablation)Cosine-LIRA · CREMA-DAUC: 0.7500 -> 0.5140Hurt
add modality-aware adaptive weighting and calibration of MIA scoresMIA on multimodal FL · CREMA-D 6 clientsAUC: 0.7500 -> 0.7802Improved
add modality-aware attack modelMIA on multimodal FL · CREMA-D 6 clientsTPR @ 1% FPR: 12.28 -> 16.67Improved
reduce number of colluding clients from 10 to 2 (ablation)proposed MIA · Balanced datasetTPR @ 1% FPR: 14.86 -> 11.86 (ASL-Gating)Hurt
reduce reference models from 6 to 0 (ablation)Likelihood-based MIA (Cosine-LiRA M) · CREMA-D-10TPR @ 1% FPR: 10.01 -> 0.00Hurt
reduce reference models from 6 to 0 (ablation)proposed MIA · CREMA-D-10TPR @ 1% FPR: 16.33 -> 15.32No clear effect
more severe overfitting (ablation)proposed MIA · CREMA-D-10AUC: .6473 -> .8576Improved
non-IID modality-quality heterogeneity across clients (ablation)MIAs generally · CREMA-D-6AUC: Cosine-LiRA (M) client 0: .8769Improved
add OGM modality balancing as defense (ablation)MFL · CREMA-Dattack AUC: 0.7798 -> 0.5697Improved
add uni-modal loss regularization as defense (ablation)MFL · CREMA-Dmodel accuracy: 55.17 -> 56.11Improved
add client-level DP with sigma 0.1 (ablation)MFL · CREMA-Dmodel accuracy: 55.17 -> 34.63Hurt

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