Jiaming Yan

Compression is an efficient way to relieve the tremendous communication overhead of federated learning (FL) systems. However, for the existing works, the information loss under compression will lead to unexpected model/gradient deviation for the FL training, significantly degrading the training performance, especially under the challenges of data heterogeneity and model obsolescence. To strike a delicate trade-off between model accuracy and traffic cost, we propose Caesar, a novel FL framework with a low-deviation compression approach. For the global model download, we design a greedy method to optimize the compression ratio for each device based on the staleness of the local model, ensuring a precise initial model for local training. Regarding the local gradient upload, we utilize the device's local data properties (i.e., sample volume and label distribution) to quantify its local gradient's importance, which then guides the determination of the gradient compression ratio. We have implemented Caesar, on two physical platforms with 40 smartphones and 80 NVIDIA Jetson devices. Extensive results show that Caesar, can reduce the traffic costs by about 25.54%þicksim37.88% when achieving the same target accuracy compared to the compression-based baselines, while incurring only a 0.68% degradation in final test accuracy relative to the full-precision communication.

Federated fine-tuning (FedFT) provides an effective paradigm for fine-tuning large language models (LLMs) in privacy-sensitive scenarios. However, practical deployment remains challenging due to the limited resources on end devices. Existing methods typically utilize parameter-efficient fine-tuning (PEFT) techniques, such as Low-Rank Adaptation (LoRA), to substantially reduce communication overhead. Nevertheless, significant memory usage for activation storage and computational demands from full backpropagation remain major barriers to efficient deployment on resource-constrained end devices. Moreover, substantial resource heterogeneity across devices results in severe synchronization bottlenecks, diminishing the overall fine-tuning efficiency. To address these issues, we propose FedQuad, a novel LoRA-based FedFT framework that adaptively adjusts the LoRA depth (the number of consecutive tunable LoRA layers from the output) according to devices’ computational power, while employing activation quantization to reduce memory overhead, thereby enabling efficient deployment on resource-constrained devices. Specifically, FedQuad first identifies the feasible and efficient combinations of LoRA depth and the number of activation quantization layers based on device-specific resource constraints. Subsequently, FedQuad employs a greedy strategy to select the optimal configurations for each device, effectively accommodating system heterogeneity. Extensive experiments demonstrate that FedQuad achieves a 1.4–5.3× convergence acceleration compared to state-of-the-art baselines when reaching target accuracy, highlighting its efficiency and deployability in resource-constrained and heterogeneous end-device environments.

Decentralized federated learning (DFL) has gained popularity for training machine learning models on massive data in edge computing, as it avoids the potential bottleneck of conventional parameter server architectures. However, the existing DFL solutions typically use deterministic topologies that struggle with both system heterogeneity and non-IID local data, resulting in high bandwidth costs and slow convergence rates. In this paper, we propose a novel mechanism called Communication-efficient Decentralized Federated Learning (CedFL) to accelerate model training. In CedFL, each worker will communicate with each of its neighbors (i.e., model exchange) according to a certain probability at each epoch, so as to reduce bandwidth consumption. To this end, we then propose an efficient algorithm to adaptively determine the optimal probability for each worker pair according to real-time system situations (e.g., data distribution and bandwidth resource). Our proposed mechanism has been extensively tested on classical models and datasets, and the results demonstrate its high effectiveness. CedFL has been shown to reduce completion time for model training by approximately 55% and improve test accuracy by 11% under the bandwidth constraint, compared to state-of-the-art solutions.

Federated Learning (FL) enables distributed clients to collaboratively train models without exposing their private data. However, it is difficult to implement efficient FL due to limited resources. Most existing works compress the transmitted gradients or prune the global model to reduce the resource cost, but leave the compressed or pruned parameters under-optimized, which degrades the training performance. To address this issue, the neural composition technique constructs size-adjustable models by composing low-rank tensors, allowing every parameter in the global model to learn the knowledge from all clients. Nevertheless, some tensors can only be optimized by a small fraction of clients, thus the global model may get insufficient training, leading to a long completion time, especially in heterogeneous edge scenarios. To this end, we enhance the neural composition technique, enabling all parameters to be fully trained. Further, we propose a lightweight FL framework, called Heroes, with enhanced neural composition and adaptive local update. A greedy-based algorithm is designed to adaptively assign the proper tensors and local update frequencies for participating clients according to their heterogeneous capabilities and resource budgets. On this basis, we further propose an extension of Heroes, termed AdaHeroes, which further improves the training performance under the statistical heterogeneity scenario based on an adaptive client selection strategy. Extensive experiments demonstrate that Heroes can reduce traffic consumption by about 72.46% and provide up to 2.76× speedup compared to the baselines. Furthermore, with the setting of statistical heterogeneity, AdaHeroes can improve the test accuracy by about 4.77% compared with Heroes and the baselines.

In edge computing (EC), federated learning (FL) enables numerous distributed devices (or workers) to collaboratively train AI models without exposing their local data. Most works of FL adopt a predefined architecture on all participating workers for model training. However, since workers’ local data distributions vary heavily in EC, the predefined architecture may not be the optimal choice for every worker. It is also unrealistic to manually design a high-performance architecture for each worker, which requires intense human expertise and effort. In order to tackle this challenge, neural architecture search (NAS) has been applied in FL to automate the architecture design process. Unfortunately, the existing federated NAS frameworks often suffer from the difficulties of system heterogeneity and resource limitation. To remedy this problem, we present a novel framework, termed Peaches , to achieve efficient searching and training in the resource-constrained EC system. Specifically, the local model of each worker is stacked by base cell and personal cell, where the base cell is shared by all workers to capture the common knowledge and the personal cell is customized for each worker to fit the local data. We determine the number of base cells, shared by all workers, according to the bandwidth budget on the parameters server. Besides, to relieve the data and system heterogeneity, we find the optimal number of personal cells for each worker based on its computing capability. In addition, we gradually prune the search space during training to mitigate the resource consumption. We evaluate the performance of Peaches through extensive experiments, and the results show that Peaches can achieve an average accuracy improvement of about 6.29% and up to 3.97× speed up compared with the baselines.

Federated learning (FL) has been widely adopted to train machine learning models over massive data in edge computing. Most works of FL employ pre-defined model architectures on all participating clients for model training. However, these pre-defined architectures may not be the optimal choice for the FL setting since manually designing a high-performance neural architecture is complicated and burdensome with intense human expertise and effort, which easily makes the model training fall into the local suboptimal solution. To this end, Neural Architecture Search (NAS) has been applied to FL to address this critical issue. Unfortunately, the search space of existing federated NAS approaches is extraordinarily large, resulting in unacceptable completion time on the resource-constrained edge clients, especially under the non-independent and identically distributed (non-IID) setting. In order to remedy this, we propose a novel framework, called FINCH, which adopts hierarchical neural architecture search to enhance federated learning. In FINCH, we first divide the clients into several clusters according to the data distribution. Then, some subnets are sampled from a pre-trained supernet and allocated to the specific client clusters for searching the optimal model architecture in parallel, so as to significantly accelerate the process of model searching and training. The extensive experimental results demonstrate the high effectiveness of our proposed framework. Specifically, FINCH can reduce the completion time by about 30.6%, and achieve an average accuracy improvement of around 9.8% compared with the baselines.