Federated learning (FL) for large language models (LLMs) offers a privacy-preserving scheme, enabling clients to collaboratively fine-tune locally deployed LLMs or smaller language models (SLMs) without exchanging raw data. While parameter-sharing methods in traditional FL models solves number of technical challenges, they still incur high communication overhead and struggle with adapting to heterogeneous model architectures. Federated distillation, a framework for mutual knowledge transfer via shared logits, typically offers lower communication overhead than parameter-sharing methods. However, transmitting logits from LLMs remains challenging for bandwidth-limited clients due to their high dimensionality. In this work, we focus on a federated LLM distillation with efficient communication overhead. To achieve this, we first propose an adaptive Top-k logit selection mechanism, dynamically sparsifying logits according to real-time communication conditions. Then to tackle the dimensional inconsistency introduced by the adaptive sparsification, we design an adaptive logits aggregation scheme, effectively alleviating the artificial and uninformative inputs introduced by conventional zero-padding methods. Finally, to enhance the distillation effect, we incorporate LoRA-adapted hidden-layer projection from LLM into the distillation loss, reducing the communication overhead further while providing richer representation. Experimental results demonstrate that our scheme achieves superior performance compared to baseline methods while effectively reducing communication overhead by approximately 50%.

Federated fine-tuning of pre-trained Large Language Models (LLMs) enables task-specific adaptation across diverse datasets while preserving data privacy. However, the large model size and heterogeneity in client resources pose significant computational and communication challenges. To address these issues, in this paper, we propose a novel Heterogeneous Adaptive Federated Low-Rank Adaptation (LoRA) fine-tuned LLM framework (HAFL). To accommodate client resource heterogeneity, we first introduce an importance-based parameter truncation scheme, which allows clients to have different LoRA ranks, and smoothed sensitivity scores are used as importance indicators. Despite its flexibility, the truncation process may cause performance degradation. To tackle this problem, we develop an importance-based parameter freezing scheme. In this approach, both the cloud server and clients maintain the same LoRA rank, while clients selectively update only the most important decomposed LoRA rank-1 matrices, keeping the rest frozen. To mitigate the information dilution caused by the zero-padding aggregation method, we propose an adaptive aggregation approach that operates at the decomposed rank-1 matrix level. Experiments on the 20 News Group classification task show that our method converges quickly with low communication size, and avoids performance degradation when distributing models to clients compared to truncation-based heterogeneous LoRA rank scheme. Additionally, our adaptive aggregation method achieves faster convergence compared to the zero-padding approach.

Over-the-air federated learning (OTA-FL) presents a promising distributed machine learning paradigm that improves the efficiency of local update aggregation by leveraging the superposition property of wireless multiple access channels (MACs). However, it faces significant security and privacy concerns that demand careful consideration. To address these threats associated with OTA-FL, we develop a secure and private over-the-air federated learning (SP-OTA-FL) framework, which can realize the secure and private aggregation for both OTA-FL with unbiased aggregation (UB-OTA-FL) and OTA-FL with biased aggregation (B-OTA-FL). In this framework, a subset of devices participate in training, while another subset functions as jammers, emitting jamming signals to enhance the security and privacy of the OTA-FL process. In particular, we measure the privacy leakage of users’ data using differential privacy (DP) and introduce an innovative application of mean squared error security (MSE-security) to evaluate the security of the OTA-FL system. We conduct convergence analyses for both convex and non-convex loss functions. Building on these analytical results, we separately formulate optimization problems for UB-OTA-FL and B-OTA-FL to enhance the learning performance of SP-OTA-FL by strategically optimizing the scheduling of training participants and jammers. The effectiveness of the proposed schemes is verified through simulations.

Federated learning (FL) has been widely investigated in academic and industrial fields to resolve the issue of data isolation in the distributed Internet of Things (IoT) while maintaining privacy. However, challenges persist in ensuring adequate privacy and security during the aggregation process. In this article, we investigate device scheduling strategies that ensure the security and privacy of wireless FL. Specifically, we measure the privacy leakage of user data using differential privacy (DP) and assess the security level of the system through the mean-square error security (MSE-security). We commence by deriving the analytical results that reveal the impact of the device scheduling on privacy and security protection, as well as on the learning process. Drawing from these analytical findings, we propose three scheduling policies that can achieve secure aggregation of wireless FL under different cases of channel noise. In particular, we formulate an integer nonlinear fractional programming problem to improve the learning performance while guaranteeing privacy and security of wireless FL. We provide an insightful solution in the closed form to the optimization problem when the model has a high dimension. For the general case, we propose a secure and private aggregation (SPA) algorithm based on the branch-and-bound (BnB) method, which can obtain the optimal solution with low complexity. The effectiveness of the proposed schemes for device selection is validated through simulations.

This paper develops an optimal design for device scheduling, alignment coefficient, and aggregation rounds within a differentially private over-the-air federated averaging (DP-OTA-FedAvg) system considering a constrained sum power budget. In DP-OTA-FedAvg, gradients are aligned using an alignment coefficient and then aggregated over the air, utilizing channel noise to ensure participant privacy. This study highlights two critical tradeoffs in aligned over-the-air federated learning (OTA-FL) systems with limited power and privacy budgets. Firstly, it reveals the tradeoff between the number of scheduled devices and the alignment coefficient. Secondly, it investigates the balance between aggregation distortion and local training error while adhering to the sum power constraint. Specifically, we measure privacy using differential privacy (DP) and perform convergence analyses for both convex and non-convex loss functions. These analyses provide insights into how device scheduling, the alignment coefficient, and the number of global aggregations affect both privacy preservation and the learning process. Building on these analytical results, we formulate an optimization problem aimed at minimizing the optimality gap of DP-OTA-FedAvg under power and privacy constraints. By specifying the number of aggregation rounds, we derive a closed-form expression describing the relationship between the alignment coefficient and the number of scheduled devices. We then tackle the problem through iterative optimization of scheduling and aggregation rounds. The effectiveness of the proposed policies is verified through simulations, and the performance advantage is particularly pronounced in scenarios where devices have poor channel conditions and limited sum-power budgets.

In this paper, we propose a device scheduling scheme for differentially private over-the-air federated learning (DP-OTA-FL) systems, referred to as S-DPOTAFL, where the privacy of the participants is guaranteed by channel noise. In S-DPOTAFL, the gradients are aligned by the alignment coefficient and aggregated via over-the-air computation (AirComp). The scheme schedules the devices with better channel conditions in the training to avoid the problem that the alignment coefficient is limited by the device with the worst channel condition in the system. We conduct the privacy and convergence analysis to theo-retically demonstrate the impact of device scheduling on privacy protection and learning performance. To improve the learning accuracy, we formulate an optimization problem with the goal to minimize the training loss subjecting to privacy and transmit power constraints. Furthermore, we present the condition that the S-DPOTAFL performs better than the DP-OTA-FL without considering device scheduling (NoS-DPOTAFL). The effectiveness of the S-DPOTAFL is validated through simulations.

To alleviate the negative impact of noise on wireless federated learning (FL), we propose a channel-weighted aggregation scheme of FL (CWA-FL), in which the parameter server (PS) makes aggregation of the gradients according to the channel conditions of devices. In the proposed scheme, the gradients are transmitted to the PS in an uncoded way through an orthogonal multiple access (OMA) channel, which can avoid the synchronization issue among devices faced by over-the-air FL. The convergence analysis of CWA-FL is conducted and the theoretical results show that the scheme can converge with the rate of$\mathcal {O} (\frac{1}{T})$. Simulation results show that the proposed scheme performs better than the equal-weighted aggregation scheme of FL (EWA-FL) and is more robust to noise.