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..
MARS: A Malignity-Aware Backdoor Defense in Federated Learning
Authors: Wei Wan, Ning Yuxuan, Zhicong Huang, Cheng Hong, Shengshan Hu, Ziqi Zhou, Yechao Zhang, Tianqing Zhu, Wanlei Zhou, Leo Yu Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that MARS can defend against SOTA backdoor attacks and significantly outperforms existing defenses. 5 Experiments 5.1 Experimental Setup 5.2 Experimental Results |
| Researcher Affiliation | Collaboration | 1 Faculty of Data Science, City University of Macau 2 School of Computing, Australian National University 3 Ant Group 4 School of Cyber Science and Engineering, Huazhong University of Science and Technology 5 School of Computer Science and Technology, Huazhong University of Science and Technology 6 College of Computing and Data Science, Nanyang Technological University 7 School of Information and Communication Technology, Griffith University |
| Pseudocode | Yes | Algorithm 1 MARS |
| Open Source Code | Yes | The codes are available at https://github.com/yunming181920/MARS. |
| Open Datasets | Yes | We evaluate the effectiveness of MARS on MNIST [16], CIFAR10 [15], and CIFAR-100 [15] datasets. |
| Dataset Splits | Yes | We consider an FL system with 100 clients, where 20 of them are designated as attackers. In each round, 20 clients are selected to participate in the FL process, with 4 of them guaranteed to be attackers. To simulate realistic non-IID distributions, we use the Dirichlet distribution with a default sampling parameter α set to 0.9. |
| Hardware Specification | Yes | We fix the random seed to ensure reproduction and conduct experiments on the NVIDIA 3090Ti. |
| Software Dependencies | No | In PyTorch, the Lipschitz constant can be easily computed using torch.svd(weight)[1].max(). (No version for PyTorch specified). we evaluated it on a pre-trained ViT model using the Hugging Face Transformers library. (No version for Hugging Face Transformers specified). |
| Experiment Setup | Yes | We consider an FL system with 100 clients, where 20 of them are designated as attackers. In each round, 20 clients are selected to participate in the FL process, with 4 of them guaranteed to be attackers. By default, MARS s hyperparameters Îș and Ï” are set to 5 and 0.03, respectively. ... For MNIST, a simple CNN is employed as the global model, while for CIFAR-10 and CIFAR-100, we use Res Net-18 [13] as the global model. |