Decentralized Federated Learning Methods for Generative AI in Edge Computing, Internet of Things (IoT) and 6G Systems

Generative AI (GAI) models, such as generative adversarial networks (GANs), variational autoencoders (VAEs) and generative pre-trained transformers (GPTs), are regarded as the key enabler of autonomous data synthesis, signal reconstruction and adaptive learning in the generative IoT and 6G systems. Facilitated by edge computing, GAI models can leverage diverse, real-world data collected from various distributed edge nodes (e.g., mobile devices, edge servers, IoT sensors, and decentralized CPU units) with the aim to train robust generative models and produce credible, trustworthy “synthetic” data.

Yet, traditional centralized approaches where all data is forwarded to a cloud for processing and storage are no longer feasible for two reasons:

i) when data is openly transmitted to, and stored on a public cloud, its privacy can be easily violated;

ii) sending/processing vast amounts of data on centralized servers overloads the communication network and computing infrastructure, leading to long delays which cannot be tolerated in real-time applications.

For these reasons, there is an urgent need for decentralized federated learning (DFL) methods for GAI. In these methods, data can be processed directly at edge nodes (i.e., without sharing) through distributed aggregation to avoid exposing sensitive data and leverage the under-used computing, communication and storage resources of the edge.

Unfortunately, although some preliminary works on federated learning (FL) methods for GAI have already been conducted, current research lack an in-depth, comprehensive framework that can facilitate efficient and feasible realization of such methods in practical systems. Specifically, the majority of existing studies on FL for GAI focus on centralized (CFL) and hierarchical FL (HFL) architectures. Such architectures require a parameter server and/or a set of controllers to coordinate edge nodes and aggregate/average their model parameters. This leads to many limitations, including a single node of failure, poor scalability, increased latency and high coordination costs. Hence, a fully decentralized FL (DFL) framework for GAI models is required to enable distributed peer-to-peer (P2P) learning across edge nodes without any assistance by the parameter server or controller. In addition, there are other critical challenges and issues which are yet to be address by the state-of-the-art, as detailed below:

1. Heterogeneous non-i.i.d. data distributions: Variations of data distributions across edge nodes can cause model drifts and inconsistencies in the generated outputs, which impacts accuracy and convergence properties of GAI models. Existing methods to address data heterogeneity and non-i.i.d. distributions are designed only for CFL and HFL systems. They are not applicable to DFL systems due to decentralization aggregation and absence of global information about edge nodes.
2. Dynamic optimization of hyperparameter settings: DFL systems for training GAI models incorporate many hyperparameters, such as learning rates, spectral radius, mixing weights, averaging step sizes, communication intervals and compression factors. Optimizing such hyperparameters is essential for enhancing performance (i.e., in terms of training accuracy and convergence properties). However, hyperparameter optimization in a practical DFL system can be challenging due to several factors, including limited, heterogeneous resources and dynamics of edge environment (i.e., dynamic topology, unreliable wireless links and unstable CPU/battery resources of edge nodes).
3. Computing and communication constraints: Unlike cloud servers, edge nodes have scarce CPU/battery resources and communicate over unreliable wireless links. Accordingly, DFL for training GAI models faces bandwidth limitations, network latency, and computational overheads. Thus, efficient methods, e.g., based on model compression, asynchronous learning, topology optimization, and edge nodes’ selection, are required to satisfy these constraints.
4. Incentive provisioning and resource allocation: Current methods for incentive provisioning and resources allocation focus on the CFL and HFL systems coordinated by the parameter server and/or controller. However, incentive provisioning in DFL is more complex than that in CFL/HFL due to decentralization and complex performance/cost trade-offs of distributed edge nodes. Hence, distributed incentive provisioning and resources allocation methods must be devised to incentivize cooperation, ensure fair participation, optimize the resources and decision-making of edge nodes in DFL for GAI models.

The objective of the proposed research is to design new DFL methods for GAI models which can address the aforementioned challenges. This includes:

1. Development of fully decentralized FL frameworks for GAI models, leveraging of gossip-based learning and distributed optimization tailored for edge computing and IoT systems.
2. Development of adaptive model fusion/aggregation strategies e.g., based on transfer learning (TL), meta-learning and reinforcement learning (RL), to handle heterogeneous non-i.i.d. data in DFL for GAI models that will allow edge nodes to intelligently update their trained GAI models based on the contextual relevance and enhance generalization across diverse data distributions of edge nodes:
3. Development of dynamic distributed algorithms for optimizing hyperparameter settings at edge nodes, e.g., based on distributed stochastic optimization methods, stochastic ADMM, Bayesian optimization, evolutionary programming, RL and meta-RL, for specific edge computing, IoT, and 6G applications.
4. Development of distributed model compression methods and self-organized agent selection algorithms for addressing computing/communication constraints of DFL for GANs, VAEs, and transformers in edge computing and IoT systems.
5. Development of novel game-theoretic methods for distributed incentive provisioning in DFL for GANs, VAEs and transformers, ensuring fair training contributions and optimal resource sharing in edge computing, IoT and 6G networks