Research map / Adversarial robustness and security / Attacks on text and graphs
Beyond Mimicking Under-Represented Emotions: Deep Data Augmentation with Emotional Subspace Constraints for EEG-Based Emotion Recognition
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 ESC-GAN, a GAN that edits well-represented EEG signals into under-represented emotional subspaces using diversity-aware and boundary-aware losses, to augment imbalanced EEG emotion recognition data.
Problem
EEG emotion recognition suffers from limited and imbalanced data; standard GANs mimic observed distributions and fail to produce reliable augmentations for under-represented minority emotions.
Approach
Instead of generating EEG from random noise, edit reference EEG from well-represented emotions toward under-represented emotional subspaces, with a diversity-aware loss that enlarges sample differences relative to feature differences and a boundary-aware loss that pushes augmented samples toward the classifier's decision boundary.
Key results
- On DEAP, ESC-GAN reaches 96.33% valence and 96.68% arousal accuracy, versus 93.52%/94.21% for the prior GAN augmentation method by Zhang et al. 2022.
- On AMIGOS under FGSM attack, accuracy rises from 66.15% without augmentation to 71.22% with ESC-GAN augmentation.
- In hard-case experiments, minority-emotion accuracy gains from augmentation can exceed 20% when only 10-50% of training samples remain in the control quadrant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ESC-GAN augmentation instead of no augmentation | EEG emotion classifier · DEAP | accuracy (valence): 94.17 -> 96.33 | Helped |
| Use ESC-GAN augmentation instead of no augmentation | EEG emotion classifier · DEAP | accuracy (arousal): 94.75 -> 96.68 | Helped |
| Use ESC-GAN instead of the strongest reported oversampling baseline | EEG emotion classifier · AMIGOS | accuracy (valence): 93.77 -> 94.40 | Helped |
| Use ESC-GAN instead of the strongest reported oversampling baseline | EEG emotion classifier · AMIGOS | accuracy (arousal): 94.42 -> 95.23 | Helped |
| Use ESC-GAN augmentation rather than no augmentation | EEG emotion classifier · SEED | accuracy: 95.22% -> 97.14 | Helped |
| Add reference EEG editing rather than generating EEG from random vectors (ablation) | ESC-GAN · DEAP | accuracy (valence): 95.07 -> 96.33 | Helped |
| Add reference EEG editing rather than generating EEG from random vectors (ablation) | ESC-GAN · DEAP | accuracy (arousal): 95.28 -> 96.68 | Helped |
| Add one-side label smoothing (ablation) | ESC-GAN · DEAP | accuracy (valence): 96.11 -> 96.33 | Helped |
| Add one-side label smoothing (ablation) | ESC-GAN · DEAP | accuracy (arousal): 96.63 -> 96.68 | Helped |
| Add diversity-aware loss (ablation) | ESC-GAN · DEAP | accuracy (valence): 95.83 -> 96.33 | Helped |
| Add diversity-aware loss (ablation) | ESC-GAN · DEAP | accuracy (arousal): 95.9 -> 96.68 | Helped |
| Add boundary-aware loss (ablation) | ESC-GAN · DEAP | accuracy (valence): 96.23 -> 96.33 | Helped |
Surprising findings
- The boundary-aware loss worsens distribution-matching metrics (MMD, FID) while improving classification accuracy, suggesting a trade-off between realism and augmentation utility.
Limitations
- Gains on the balanced SEED dataset are smaller and ESC-GAN ranks just behind GANSER there.
- Adversarial robustness evaluated only on AMIGOS with arousal and FGSM/PGD/DeepFool at a single small epsilon.
Experimental setup
Datasets: DEAP AMIGOS SEED
Benchmarks: DEAP AMIGOS SEED
Metrics: accuracy Kernel MMD Inception Score FID
Baselines: SMOTE ADASYN Zhang et al. 2022 GAN augmentation Luo and Lu 2018 Liu et al. 2023 Topic and Russo 2021 Feng et al. 2022 Shen et al. 2020
Builds on: GAN mixup FGSM
Closest papers by meaning
- EMOD: A Unified EEG Emotion Representation Framework Leveraging V-A Guided Contrastive Learning (AAAI 2026 · similarity 0.88)
- VBH-GNN: Variational Bayesian Heterogeneous Graph Neural Networks for Cross-subject Emotion Recognition (ICLR 2024 · similarity 0.86)
- DMMR: Cross-Subject Domain Generalization for EEG-Based Emotion Recognition via Denoising Mixed Mutual Reconstruction (AAAI 2024 · similarity 0.86)
- Multi-dataset Joint Pre-training of Emotional EEG Enables Generalizable Affective Computing (NeurIPS 2025 · similarity 0.85)
- EEVR: A Dataset of Paired Physiological Signals and Textual Descriptions for Joint Emotion Representation Learning (NeurIPS 2024 · similarity 0.84)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.83)
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.