Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Enhancing Privacy in Multimodal Federated Learning with Information Theory
Authors: Tianzhe Xiao, Yichen Li, Yining Qi, YI LIU, wangshi.ww, Haozhao Wang, Yi Wang, Ruixuan Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments have proven that our method can bring more balanced and comprehensive protection at an acceptable cost. 5 Experiments We conduct MMFL experiments using both synthesized and native multimodal datasets. For image-text modality studies, we employ CIFAR-10/100 with text descriptions generated from image labels, following standard GIA experimental protocols. Additionally, we validate our method on authentic multimodal benchmarks: Hateful-Memes for social media content analysis and Crisis MMD for disaster response. |
| Researcher Affiliation | Collaboration | Tianzhe Xiao1, Yichen Li1, Yining Qi1, Yi Liu2, Wei Wang2, Haozhao Wang1 , Yi Wang2, Ruixuan Li1 1School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China 2Chongqing Ant Consumer Finance Co., Ltd, Ant Group, Chongqing, China EMAIL |
| Pseudocode | Yes | Algorithm 1: Sec-MMFL |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Crisis MMD dataset access is restricted by proprietary agreements. Synthetic CIFAR-text datasets can be regenerated using our described methodology. |
| Open Datasets | Yes | We conduct MMFL experiments using both synthesized and native multimodal datasets. For image-text modality studies, we employ CIFAR-10/100 with text descriptions generated from image labels, following standard GIA experimental protocols. Additionally, we validate our method on authentic multimodal benchmarks: Hateful-Memes for social media content analysis and Crisis MMD for disaster response. ... CIFAR datasets (CC-BY 4.0) and Hateful Memes (CC-BY-NC 2.0) licenses acknowledged in References. Crisis MMD usage complies with original data sharing agreements. |
| Dataset Splits | No | The paper mentions using specific datasets (CIFAR-10/100, Hateful-Memes, Crisis MMD) but does not provide specific percentages, sample counts, or explicit splitting methodologies for training, validation, and test sets. It refers to "standard GIA experimental protocols" but without specifying the splits used within those protocols. |
| Hardware Specification | Yes | We run all experiments on Intel Xeon Gold 6133 CPU, RTX4090 GPU. |
| Software Dependencies | No | The resulting σi values are applied to each modality s Privacy Engine in Opacus, enabling efficient privacy-preserving training with minimal utility loss. However, no version numbers for Opacus or any other software components are provided. |
| Experiment Setup | Yes | We take λ as 1e-2 chosen via grid search, δ as 1e-5 and clipping norm as 1.0. The batch size during training is set to 128. The learning rate η is set to 1e-3 and the training is conducted for 200 rounds. |