Penghui Yang, Cunxiao Du, Fengzhuo Zhang, Haonan Wang, Tianyu Pang, Chao Du, Bo An. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In Federated Learning (FL), the decentralized nature of client training introduces vulnerabilities, notably backdoor attacks. Prevailing anomaly detection approaches typically perform binary classification, dividing clients into trusted and untrusted groups. However, these methods face two critical challenges: the insider threat, where malicious clients concealed within the trusted group compromise the global model, and the benign exclusion, where legitimate contributions from benign clients are mistakenly classified as untrusted and disregarded. These issues weaken both the robustness and fairness of FL systems, exposing inherent defense vulnerabilities. To address these challenges, we propose FedDLAD, a Federated Learning Dual-Layer Anomaly Detection framework designed to enhance resilience against backdoor attacks. The framework leverages the Connectivity-Based Outlier Factor (COF) module to perform a robust initial classification of clients by analyzing structural data connectivity. The Interquartile Range (IQR) module further reinforces this by mitigating the insider threat through the removal of residual malicious influences within the trusted group. Furthermore, the Pardon module dynamically reintegrates misclassified benign clients from the untrusted group, thereby preserving their valuable contributions and addressing the benign exclusion. We conduct extensive evaluations of FedDLAD against state-of-the-art defenses on real-world datasets, demonstrating its superior ability to reduce backdoor attack success rates while maintaining robust model performance. Code is available at: https://github.com/dingbinb/FedDLAD.

Traditional knowledge distillation focuses on aligning the student's predicted probabilities with both ground-truth labels and the teacher's predicted probabilities. However, the transition to predicted probabilities from logits would obscure certain indispensable information. To address this issue, it is intuitive to additionally introduce a logit-level loss function as a supplement to the widely used probability-level loss function, for exploiting the latent information of logits. Unfortunately, we empirically find that the amalgamation of the newly introduced logit-level loss and the previous probability-level loss will lead to performance degeneration, even trailing behind the performance of employing either loss in isolation. We attribute this phenomenon to the collapse of the classification head, which is verified by our theoretical analysis based on the neural collapse theory. Specifically, the gradients of the two loss functions exhibit contradictions in the linear classifier yet display no such conflict within the backbone. Drawing from the theoretical analysis, we propose a novel method called dual-head knowledge distillation, which partitions the linear classifier into two classification heads responsible for different losses, thereby preserving the beneficial effects of both losses on the backbone while eliminating adverse influences on the classification head. Extensive experiments validate that our method can effectively exploit the information inside the logits and achieve superior performance against state-of-the-art counterparts

Federated learning (FL) enables multiple clients to collaboratively train machine learning models under the coordination of a central server, while maintaining privacy. However, the server cannot directly monitor the local training processes, leaving room for malicious clients to introduce backdoors into the model. Research has shown that backdoor attacks exploit specific neurons that are activated only by malicious inputs, remaining dormant with clean data. Building on this insight, we propose a novel defense method called Flipping Weight Updates of Low-Activation Input Neurons (FLAIN) to counter backdoor attacks in FL. Specifically, upon the completion of global training, we use an auxiliary dataset to identify low-activation input neurons and iteratively flip their associated weight updates. This flipping process continues while progressively raising the threshold for low-activation neurons, until the model's performance on the auxiliary data begins to degrade significantly. Extensive experiments demonstrate that FLAIN effectively reduces the success rate of backdoor attacks across a variety of scenarios,including Non-IID data distributions and high malicious client ratios (MCR), while maintaining minimal impact on the performance of clean data. The source code is available at: FLAIN.

Noisy label learning (NLL) in open-world scenarios poses a novel challenge due to the presence of noisy data from both known and unknown classes. Most existing methods operate under the closed-set assumption, rendering them vulnerable to open-set noise, which significantly degrades their performance. While some approaches attempt to mitigate the impact of open-set examples, they struggle to learn effective discriminative representations for them, leading to unsatisfactory recognition performance. To address these issues, we propose a unified Open Adapter (OpenAda) that identifies open-set noise from both data-centric and learning-based perspectives, and can be easily integrated into mainstream NLL methods to improve their performance and robustness. Specifically, the data-centric part leverages label clusterability to sequentially identify basic clean and basic open-set examples both with high neighbor agreement. The learning-based part integrates one-vs-all classifiers with a progressive open disambiguation strategy to learn a reliable “inlier vs. outlier” boundary for each class. This enables the model to detect challenging open-set examples that partially overlap in the representation space with closed-set ones. Extensive experiments on synthetic and real-world datasets validate the superiority of our approach. Notably, with minor modifications, DivideMix with OpenAda achieves performance improvements of 9.31% and 18.26% on the open-world CIFAR-80 dataset under 80% symmetric noise and 40% asymmetric noise. The code is available athttps://github.com/chenchenzong/OpenAda.

Existing knowledge distillation methods typically work by imparting the knowledge of output logits or intermediate feature maps from the teacher network to the student network, which is very successful in multi-class single-label learning. However, these methods can hardly be extended to the multi-label learning scenario, where each instance is associated with multiple semantic labels, because the prediction probabilities do not sum to one and feature maps of the whole example may ignore minor classes in such a scenario. In this paper, we propose a novel multi-label knowledge distillation method. On one hand, it exploits the informative semantic knowledge from the logits by dividing the multi-label learning problem into a set of binary classification problems; on the other hand, it enhances the distinctiveness of the learned feature representations by leveraging the structural information of label-wise embeddings. Experimental results on multiple benchmark datasets validate that the proposed method can avoid knowledge counteraction among labels, thus achieving superior performance against diverse comparing methods. Our code is available at: https://github.com/penghui-yang/L2D.