CVPR Findings2026
FLToM: Robust Federated Learning with Theory-of-Mind Structure
Tianshu Xiao, Liu Yang, Sichang Guo, Qilong Wang, Qinghua Hu
Abstract
The security of Federated Learning (FL) is critically challenged by strategic adversaries employing coordinated and blended attack methodologies. While many defenses offer theoretical guarantees, they often rely on potentially fallible similarity metrics, rendering them less effective against such attacks. We argue that a robust framework must provide two distinct capabilities: 1) effective detection and filtering to weaken adversarial influence, and 2) attack classification to guide subsequent defense policies. This paper introduces a novel approach that, to the best of our knowledge, is the first to incorporate attack classification into the defense process by leveraging the Theory of Mind (ToM). We propose FLToM, a framework that learns to form unified, abstract models of both malicious and benign clients from their behavior, inspired by the machine theory of mind. FLToM implements a ToM hierarchy to infer latent intentions and classifies clients into distinct categories. By integrating reasoning from communication, history, and a proxy dataset, FLToM enables robust filtering while providing client-specific ToM embedding for advanced defense policies. Extensive experiments demonstrate that FLToM significantly outperforms state-of-the-art methods, offering a novel framework for applying the broad and growing literature of FL defenses to real-world applications.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on17
- FLDetector: Defending Federated Learning Against Model Poisoning Attacks via Detecting Malicious ClientsZaixi Zhang, Xiaoyu Cao, Jinyuan Jia, Neil Zhenqiang GongKDD 2022 · 293 citations
- Byzantine-Robust Learning on Heterogeneous Datasets via BucketingSai Praneeth Karimireddy, Lie He, Martin JaggiICLR 2022 · 192 citations
- Sageflow: Robust Federated Learning against Both Stragglers and AdversariesJungwuk Park, Dong-Jun Han, Minseok Choi, Jaekyun MoonNeurIPS 2021 · 149 citations
- ToM2C: Target-oriented Multi-agent Communication and Cooperation with Theory of MindYuanfei Wang, Fangwei Zhong, Jing Xu, Yizhou WangICLR 2022 · 103 citations
- FedInv: Byzantine-Robust Federated Learning by Inversing Local Model UpdatesBo Zhao, Peng Sun, Tao Wang, Keyu JiangAAAI 2022 · 82 citations
Related papers
- Safe-FedLLM: Delving into the Safety of Federated Large Language ModelsMingxiang Tao, Yu Tian, Wenxuan Tu, Yue Yang et al.ACL 2026 · 2 citations
- Tracing Back the Malicious Clients in Poisoning Attacks to Federated LearningYuqi Jia, Minghong Fang, Hongbin Liu, Jinghuai Zhang et al.NeurIPS 2025 · 8 citations
- Refiner: Data Refining against Gradient Leakage Attacks in Federated LearningMingyuan Fan, Cen Chen, Chengyu Wang, Xiaodan Li et al.USENIX Security 2025
- FedREDefense: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction ErrorYueqi Xie, Minghong Fang, Neil Zhenqiang GongICML 2024 · 32 citations
- RECESS Vaccine for Federated Learning: Proactive Defense Against Model Poisoning AttacksHaonan Yan, Wenjing Zhang, Qian Chen, Xiaoguang Li et al.NeurIPS 2023 · 26 citations
