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

BrainGuard: Privacy-Preserving Multisubject Image Reconstructions from Brain Activities

AAAI 2025 · Federated learning and privacy › Privacy attacks and secure learning · approach: Visual privacy and face de-identification · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BrainGuard, a privacy-preserving collaborative global-local framework that trains per-subject fMRI models with a shared global model and hybrid layer-wise synchronization to reconstruct images from multisubject fMRI on NSD.

Problem

Multisubject fMRI-to-image reconstruction pools all subjects' brain data, raising privacy concerns, and simple subject-specific layers (e.g., ridge regression) fail to handle individual variability.

Approach

A global-local architecture where each subject has a local model trained only on its own fMRI and a global model aggregates parameters (FedAvg-style with EMA), with a hybrid synchronization strategy: retain foundational layers, copy intermediate layers from global, and softly fuse advanced layers via learned Dynamic Fusion Learner weights.

Key results

• BrainGuard reports PixCorr 0.313 and SSIM 0.330, matching or exceeding the best prior methods on low-level metrics (MindEye, Psychometry, Brain-Diffuser).
• On high-level metrics BrainGuard reaches Inception 96.1%, CLIP 96.4%, EffNet-B 0.624 and SwAV 0.353, beating MindEye and Psychometry.
• Ablation shows applying all three synchronization parts (foundational, intermediate, advanced) gives the best results; the no-synchronization baseline reaches only PixCorr 0.163 and CLIP 86.1%.

Measured effects reported in the paper

Surprising findings

• The ablation indicates the 8-layer DFL configuration is optimal, and both fewer and more aggregated layers perform worse.

Limitations

• Only designed for fMRI; does not address MEG or EEG.
• Focuses on visual cortex signals only, not other brain regions.
• Evaluated only on four NSD subjects with a fixed voxel mask.

Experimental setup

Datasets: Natural Scenes Dataset (NSD) MS-COCO

Benchmarks: NSD test set

Metrics: PixCorr SSIM AlexNet(2) AlexNet(5) Inception CLIP EffNet-B SwAV

Baselines: Mind-Reader Mind-Vis Takagi et al. Gu et al. Brain-Diffuser MindEye MindBridge Psychometry

Builds on: FedAvg FedALA MindBridge Versatile Diffusion SoftCLIP

Closest papers by meaning

• MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data (ICML 2024 · similarity 0.86)
• MindBridge: A Cross-Subject Brain Decoding Framework (CVPR 2024 · similarity 0.85)
• Psychometry: An Omnifit Model for Image Reconstruction from Human Brain Activity (CVPR 2024 · similarity 0.85)
• Toward Generalizing Visual Brain Decoding to Unseen Subjects (ICLR 2025 · similarity 0.85)
• See Through Their Minds: Learning Transferable Brain Decoding Models from Cross-Subject fMRI (AAAI 2025 · similarity 0.84)
• MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding (NeurIPS 2025 · similarity 0.84)