Federated Learning (FL), a privacy-preserving decentralized machine learning framework, has been shown to be vulnerable to backdoor attacks. Current research primarily focuses on the Single-Label Backdoor Attack (SBA), wherein adversaries share a consistent target. However, a critical fact is overlooked: adversaries may be non-cooperative, have distinct targets, and operate independently, which exhibits a more practical scenario called Multi-Label Backdoor Attack (MBA). Unfortunately, prior works are ineffective in the MBA scenario since non-cooperative attackers exclude each other. In this work, we conduct an in-depth investigation to uncover the inherent constraints of the exclusion: similar backdoor mappings are constructed for different targets, resulting in conflicts among backdoor functions. To address this limitation, we propose Mirage, the first non-cooperative MBA strategy in FL that allows attackers to inject effective and persistent backdoors into the global model without collusion by constructing in-distribution (ID) backdoor mapping. Specifically, we introduce an adversarial adaptation method to bridge the backdoor features and the target distribution in an ID manner. Additionally, we further leverage a constrained optimization method to ensure the ID mapping survives in the global training dynamics. Extensive evaluations demonstrate that Mirage outperforms various state-of-the-art attacks and bypasses existing defenses, achieving an average ASR greater than 97% and maintaining over 90% after 900 rounds. This work aims to alert researchers to this potential threat and inspire the design of effective defense mechanisms. Code has been made open-source.

Federated Learning (FL) typically involves processing redundant data on resource-constrained edge devices, resulting in prolonged training time. A promising strategy to accelerate FL is on-device data sampling. SOTA methods generally select a subset based on sample importance before each local epoch. However, these methods, which employ mini-batch gradient descent on the static subset, suffer from outdated sample importance, leading to suboptimal sampling efficiency. Furthermore, these sampling methods fail to address the biased gradient expectation introduced by importance sampling, further degrading model accuracy. In this paper, we propose FedTT, a novel framework designed to accelerate FL with improved accuracy through pipelined intra-batch data sampling and model training. Specifically, at the algorithm level, to enable real-time data sampling with unbiased model updates, FedTT performs on-device sampling from the fixed-size mini-batch at each iteration and applies gradient correction to the variable-size sampled micro-batch. At the system level, to further accelerate FL training, FedTT implements a well-designed parallelism and synchronization mechanism that enables pipelined execution of intra-batch data sampling and model training on CPU-GPU architectures. Finally, we conduct extensive real-world experiments and simulations to demonstrate the effectiveness and universal adaptability of FedTT. Compared to the SOTAs, our evaluation on four datasets shows that FedTT improves time-to-accuracy performance by 1.23 × ∼ 2.51 × and model accuracy by 0.39% ∼ 11.24%, with real-world experiments further validating its real-world effectiveness by achieving a 1.75 × speedup and a 9.86% accuracy improvement. Integrated with various FL algorithms and importance criteria, FedTT consistently delivers performance gains. Code has been open sourced.

Recent years have witnessed increasing privacy concerns towards machine learning. To protect privacy in machine learning, federated learning has been proposed as a decentralized privacy-preserving framework where clients upload the parameters rather than private data. However, training a fair federated learning model in heterogeneous environments is still challenging. First, heterogeneous data distributions lead the global model fail to show high accuracy on all distributions. Second, the federated learning training process exposes and exacerbates potential biases in heterogeneous training data. Third, the local bias of each client can be propagated through parameter sharing, biasing the global model. In this work, we propose a two-stage fairness-aware federated learning framework (HeteroFair) to achieve fairness under heterogeneous data distributions. Initially, we introduce the fairness constraint to the loss function and propose a local adaptive weighting algorithm to adjust the proportion of the fairness constraint, achieving fair training in heterogeneous environments. Then, we present a fairness-aware aggregation reweighting algorithm that reduces the mismatch between local and global fairness to achieve fair federated learning. Extensive evaluation results demonstrate the effectiveness of our proposed framework in achieving fairness and high accuracy under het-eroaeneous data distributions.