Robustness without evidence: label-typical visual priors drive multimodal sentiment-analysis robustness under covert acoustic degradation across centralized and federated training
Code, frozen results, preregistration documents and decision log for the v8.2 manuscript submitted to Scientific Reports. The archive contains: (1) code/ - all analysis and figure scripts, including the CH-SIMS, IEMOCAP and MOSEI feature builders and the audit pipeline; (2) results/ - frozen JSON results and Markdown reports referenced by every number in the text, including the cross-corpus gap analyses (results/iemocap, results/mosei) and the federated/D-privacy results (results/s6) (cached features excluded; they are derivable from the public corpora CH-SIMS and CREMA-D); (3) preregistration/ - the original preregistration and three amendments with SHA-256 hashes; (4) provenance/decisions.md - the complete decision log D1-D70; (5) manuscript/ - the submission draft, supplementary information, cover letter and references; (6) figures/ - Figures 1-5 (PDF).
Authors
- Jun Zhao
Institutions
- Chengdu University of Information Technology (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24365247
- Primary Topic
- Emotion and Mood Recognition
- Type
- article
- Field-Weighted Citation Impact
- 0.00