A wide-area network federated learning algorithm based on asynchronous decoupling and elastic sharding
Patent Information
- Application Number
- CN202611015740.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-10-13
AI Technical Summary
[0005]现有技术在广域网（WAN）实体环境下面临多重难以克服的技术缺陷
[0031]本方法通过引入虚拟时空坐标系进行弹性分片，削弱了广域网延迟波动带来的影响。且通过构建数据与共识解耦的异步DAG流水线，打破了传统同步拜占庭共识的扩展瓶颈。并结合数据可用性抽样协议，大幅节省了跨分片交互的通信开销。实验表明，该算法不仅在吞吐量、延迟和带宽缩减上取得了显著成效，在面临高比例恶意攻击时也表现出了卓越的系统鲁棒性，具有较高的实际工程部署价值。
Smart Images
Figure CN122905097A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a federated learning algorithm, and more particularly to a wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding. Background Technology
[0002] With the development of edge computing and the Internet of Things (IoT), traditional federated learning frameworks based on a single central cloud are prone to communication congestion and single points of failure when faced with massive device access. Layered or sharded federated learning effectively alleviates the bandwidth pressure of global broadcast by introducing edge nodes or dividing the network into multiple sub-networks. Existing research has explored the underlying consensus of blockchain federated learning extensively. For example, some schemes use verification committee mechanisms to alleviate block generation delays or optimize the consensus process through static node clustering. However, in heterogeneous latency environments of wide area networks (WANs), existing random sharding or static clustering often struggles to adapt to dynamic changes in network topology, easily leading to excessively high intra-shard communication latency. Furthermore, traditional consensus mechanisms such as Practical Byzantine Fault Tolerance (PBFT) generally employ a tightly coupled synchronous blocking mode. As the model size continues to expand, the bandwidth waste and synchronization waiting problems caused by the frequent and direct broadcasting of large-volume parameters in the network have not been completely resolved.
[0003] To overcome the throughput bottleneck of traditional chain structures and synchronous consensus, Directed Acyclic Graphs (DAGs) are widely used in asynchronous consensus mechanisms. In the standard DAG model, nodes encapsulate transactions or proofs as vertices, forming partial order relationships by referencing predecessor vertices through hashes. This mechanism allows nodes to generate vertices concurrently without blocking, avoiding leader election and block generation waiting in traditional blockchains, thus significantly improving network bandwidth utilization. However, the local model gradients of federated learning can easily reach tens to hundreds of megabytes. Directly embedding such large model parameters into DAG vertices for concurrent broadcasting across the entire network not only fails to leverage the high throughput advantage of DAGs but also quickly exhausts physical channel resources, triggering severe network storms and channel conflicts. Therefore, finding a balance between the high concurrency characteristics of DAGs and the massive physical data transmission is a pressing technical challenge.
[0004] Erasure coding (EC), such as Reed-Solomon (RS) coding, has traditionally been an effective means of addressing single points of failure in distributed storage systems. Encoding divides the original data into equal parts using polynomial interpolation. Data blocks, and calculate Each check block is combined into Each fragment is of equal length. In an unreliable network environment, the receiver only needs to... Data can be reconstructed without loss using only a few valid fragments. In recent years, Data Availability (DA) theory has received significant attention in blockchain scaling solutions. Introducing it into federated learning and combining it with Merkle trees to construct cryptographic commitments for large-volume global model data allows for high-probability confirmation of the underlying model's true availability through cross-shard Data Availability Sampling Verification (DAS) without requiring nodes to download all parameters. This not only effectively avoids data withholding attacks launched by malicious nodes but also drastically reduces the originally linearly increasing cross-shard communication overhead to the logarithmic level, providing solid theoretical support for overcoming system scaling bottlenecks.
[0005] Existing technologies face several insurmountable technical drawbacks in wide area network (WAN) environments. On one hand, existing sharding mechanisms often rely on static clustering or random sharding, failing to perceive the actual heterogeneous latency of underlying physical nodes. This easily forces physically distant nodes into the same shard, resulting in abnormally high intra-shard physical communication latency. On the other hand, traditional synchronous Byzantine consensus, when dealing with ever-expanding model parameters, frequently broadcasting massive amounts of physical data across the network quickly exhausts physical channel bandwidth and causes long-term idle node computing resources. Furthermore, directly utilizing the non-cyclic properties of Directed Acyclic Graphs (DAGs) to broadcast large-volume models results in a massive influx of physical data packets instantly filling the channel, failing to improve throughput and instead triggering severe physical network storms and channel conflicts. Moreover, the requirement for verification nodes to download all underlying physical parameters during cross-shard interaction causes cross-shard data transmission to grow linearly, directly offsetting the bandwidth reduction advantage of the sharding architecture and becoming a physical bottleneck limiting system scalability. To overcome these physical network congestion and resource limitations, this invention proposes a WAN federated learning algorithm based on asynchronous decoupling and elastic sharding. Summary of the Invention
[0006] This invention designs and develops a wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding, which can overcome the physical topology rigidity caused by vehicle movement, avoid congestion and collapse of the vehicle wireless channel, and save the physical communication overhead of cross-regional interaction in vehicle networks.
[0007] The technical solution provided by this invention is as follows:
[0008] A wide-area network federated learning algorithm based on asynchronous decoupling and elastic sharding includes:
[0009] Step 1: Measure the round-trip time of physical nodes based on the mapping algorithm and map it to virtual space coordinates. Divide all network nodes into multiple high-cohesion shards based on coordinate distance.
[0010] Step 2: Train the local model based on all nodes in the network, generate local gradients, slice the gradients using erasure coding, construct a Merkle tree, spread the fragments within the slices, collect a quorum of verification signatures, and generate a lightweight data availability proof.
[0011] Representative nodes from each shard asynchronously broadcast and topologically sort the DAG vertices that encapsulate the data availability proof, outputting a globally strict order.
[0012] Step 3: Randomly request a small number of data fragments and Merkel proofs from the target shard, perform cross-shard sampling verification, merge the updates that pass the cross-shard verification into the global model, discard and mark malicious nodes if the process fails; repeat steps 2 and 3 until the global model is output.
[0013] Preferably, step one includes:
[0014] Based on the virtual spatiotemporal coordinate system, each physical node in the entire network iteratively updates its own coordinates along the vector direction based on the error between the actual delay and the predicted delay;
[0015] During the unconverged virtual coordinate phase, each node in the wide area network federated learning system randomly selects other nodes to probe its actual round-trip communication delay and calculates its Euclidean distance in the virtual space, using it as the basis for adaptively iteratively updating the virtual coordinates of the node.
[0016] Once all node coordinates converge, the spatial coordinates are extracted, and the entire network of nodes is divided into K highly cohesive fragment sets based on the K-means clustering algorithm.
[0017] For each sub-partition, find the node closest to its centroid as the representative node;
[0018] Outputs shards and representative nodes used to accelerate consensus.
[0019] Preferably, step two includes:
[0020] Based on erasure coding technology, ordinary nodes slice and encode large-volume local gradient models.
[0021] Construct a Merkle tree based on slice fragments and extract the global identifier root hash;
[0022] Ordinary nodes continuously send fragments and Merkel proofs to random nodes within the shard. When a valid signature is received from a verification node, it is added to the local signature set.
[0023] When the number of collected signatures reaches the statutory fault tolerance threshold Q, ordinary nodes use the BLS algorithm to merge and aggregate the signature set, assemble it with the root hash and publisher information into an availability proof, and return it to the representative node of this shard, thus completing the decoupling of physical data transmission and consensus ordering.
[0024] Preferably, step three includes:
[0025] The availability proof is used as the proof to be verified. The root hash and aggregate signature are extracted from the proof. Based on the aggregate public key of the target shard, the legality of the signature is initially verified by the BLS signature verification algorithm.
[0026] Under the condition that the signature is valid, the verification node only randomly requests a specified number of data fragments and their Merkel proof paths from the target shard.
[0027] The verification node traverses the sampled fragment set and verifies the cryptographic validity of its Merkel proof one by one.
[0028] If any fragment is tampered with or cannot be aligned with the root hash, the verification will be immediately rejected and an error will be returned.
[0029] If all sampled fragments pass the verification, it is highly probable that the underlying full model data is usable and secure.
[0030] The beneficial effects of this invention are as follows:
[0031] This method mitigates the impact of WAN latency fluctuations by introducing a virtual spatiotemporal coordinate system for elastic sharding. Furthermore, by constructing an asynchronous DAG pipeline that decouples data and consensus, it overcomes the scalability bottleneck of traditional synchronous Byzantine consensus. Combined with a data availability sampling protocol, it significantly reduces communication overhead for cross-shard interactions. Experiments show that this algorithm not only achieves significant results in throughput, latency, and bandwidth reduction, but also demonstrates excellent system robustness against high-proportion malicious attacks, making it highly valuable for practical engineering deployment.
[0032] Vehicle Nodes (OBUs) and Roadside Units (RSUs) dynamically cluster vehicles and RSUs with similar physical communication delays into highly cohesive segments by measuring round-trip physical delay (RTT) in real time and iteratively updating virtual spatiotemporal coordinates. This effectively overcomes the physical topology rigidity caused by vehicle movement.
[0033] When dealing with large-volume autonomous driving models, vehicle nodes use erasure coding to slice local model parameters and spread them only among neighboring physical vehicle fleets or RSUs. Meanwhile, the RSUs, as representative nodes, asynchronously broadcast a very small volume of data availability proofs (CoA) on the directed acyclic graph (DAG) network. This completely decouples the transmission of large entity models from the ordering of consensus metadata, avoiding congestion and collapse of the vehicle wireless channel.
[0034] Cross-regional verification nodes (RSUs in adjacent regions) do not need to download hundreds of megabytes of full autonomous driving model parameters from the target region. They only need to randomly request a very small number of physical data fragments and Merkel proofs to confirm the real availability of the underlying global model data with a high probability, which greatly saves the physical communication overhead of cross-regional interaction in vehicle-to-everything (V2X) communication. Attached Figure Description
[0035] Figure 1(a) is a schematic diagram comparing the system throughput of the consensus protocol in the MNIST dataset described in this invention.
[0036] Figure 1(b) is a schematic diagram comparing the system throughput of various consensus protocols in the Fashion-MNIST dataset described in this invention.
[0037] Figure 1(c) is a schematic diagram comparing the system throughput of various consensus protocols in the CIFAR-10 dataset described in this invention.
[0038] Figure 1(d) is a schematic diagram comparing the system throughput of various consensus protocols in the CIFAR-100 dataset described in this invention.
[0039] Figure 2(a) is a schematic diagram comparing the consensus latency of various consensus protocols in the MNIST dataset described in this invention.
[0040] Figure 2(b) is a schematic diagram comparing the consensus latency of the consensus protocol in the Fashion-MNIST dataset described in this invention.
[0041] Figure 2(c) is a schematic diagram comparing the consensus latency of the CIFAR-10 dataset consensus protocol described in this invention.
[0042] Figure 2(d) is a schematic diagram comparing the consensus latency of the CIFAR-100 dataset consensus protocol described in this invention.
[0043] Figure 3(a) is a schematic diagram showing the evolution of the total overhead of cross-slice communication across the entire network in the CIFAR-10 dataset as a function of node size.
[0044] Figure 3(b) is a schematic diagram showing the evolution of the total overhead of cross-slice communication across the entire network in the CIFAR-100 dataset as a function of node size.
[0045] Figure 4(a) is a schematic diagram comparing the global model accuracy under the 20% Byzantine malicious node ratio described in this invention.
[0046] Figure 4(b) is a schematic diagram comparing the global model accuracy under the 30% Byzantine malicious node ratio described in this invention.
[0047] Figure 5 This is a flowchart of the wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding described in this invention. Detailed Implementation
[0048] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.
[0049] As shown in Figures 1-5, this invention provides a wide-area network federated learning algorithm based on asynchronous decoupling and elastic sharding. The Asynchronous Decoupling and ElasticSharding Federated Learning (ADES-FL) algorithm performs spatiotemporal optimization and elastic sharding based on network mapping. It improves system throughput based on an asynchronous DAG architecture that decouples data and consensus, and ensures model security through a cross-shard parallel sampling verification mechanism. The algorithm includes:
[0050] Step 1: Measure the round-trip time of physical nodes based on the mapping algorithm and map it to virtual space coordinates. Divide all network nodes into multiple high-cohesion shards based on coordinate distance.
[0051] Among them, physical nodes are distributed terminal devices or edge servers that actually participate in federated learning in the wide area network;
[0052] Step 2: Train the local model based on all nodes in the network, generate local gradients, slice the gradients using erasure coding, construct a Merkle tree, spread the fragments within the slices, collect a quorum of verification signatures, and generate a lightweight data availability proof.
[0053] The network nodes refer to all participating nodes in the federated learning network, including ordinary nodes and representative nodes.
[0054] Representative nodes from each shard asynchronously broadcast and topologically sort the DAG vertices that encapsulate the data availability proof, outputting a globally strict order.
[0055] Step 3: Randomly request a small number of data fragments and Merkel proofs from the target shard, perform cross-shard sampling verification, merge the updates that pass the cross-shard verification into the global model, discard and mark malicious nodes if the process fails; repeat steps 2 and 3 until the global model is output.
[0056] The output global model is used to complete specific distributed machine learning tasks, protect the data privacy of each node, break down data silos in wide area networks, and obtain a joint generalization model with high accuracy.
[0057] Table 1. Symbol Explanation
[0058]
[0059] In the initialization phase of federated learning, the ADES-FL algorithm first uses Algorithm 2 to perform spatiotemporal mapping and complete node sharding (line 1). Then, it enters a multi-round federated consensus iteration cycle (line 2). In each iteration, all nodes in the network use local data to perform local model training in parallel (line 4), and use Algorithm 3 to generate lightweight data availability proofs from local gradients (line 5). Subsequently, it asynchronously and sequentially outputs a globally strictly ordered sequence of proofs (line 6). The system then traverses the ordered target proofs (line 8), using Algorithm 4 to perform cross-shard sampling verification (line 9). If the verification passes, it updates and merges the proof into the global model parameters (line 10). If the verification fails, the update is discarded and the malicious shard that published the proof is marked (lines 11-13). This process is repeated until the total consensus rounds are reached, finally outputting the global model (line 16). The specific algorithm is as follows:
[0060]
[0061] This algorithm is a federated learning method based on elastic sharding. While ensuring data privacy and security, it leverages the advantages of sharding to allow parallel local training across multiple highly cohesive shards. The system utilizes representative nodes to coordinate the ordering and cross-shard interactions of asynchronous Directed Acyclic Graphs (DAGs), while ordinary nodes are responsible for local model training and parameter propagation on their local data. In this way, the algorithm can accelerate the model training process by utilizing distributed computing resources while protecting user data privacy.
[0062] Perform flexible sharding and select representative nodes
[0063] To optimize communication latency in wide area networks, this invention, as a preferred approach, introduces a virtual spatiotemporal coordinate system. Nodes iteratively update their coordinates along a vector direction based on the error between the actual and predicted latency. After coordinate convergence, a clustering algorithm is used to divide the nodes into K fragments, and the node closest to the centroid is selected as the representative node to coordinate, ensuring that nodes within a fragment can complete pre-consensus at a microsecond speed, as detailed below:
[0064]
[0065] Algorithm 2 describes a process for elastic sharding and representative node selection during the network topology construction phase. During the phase where virtual coordinates are not yet converged (line 1), each node in the system randomly selects other nodes to probe its actual round-trip communication delay (lines 3-4) and calculates its Euclidean distance in virtual space (line 5). Based on this, the virtual coordinates of the nodes are adaptively iteratively updated (line 6). Once the coordinates of all nodes converge, these spatial coordinate features are extracted (line 9), and the entire network is divided into K highly cohesive shard sets using the K-means clustering algorithm (line 10). For each partitioned sub-shard, the node closest to the centroid of that shard is selected as the representative node (lines 11-13), ultimately outputting the shard and representative node set used to accelerate consensus.
[0066] Data decoupling and asynchronous sequencing
[0067] Traditional consensus protocols typically package large amounts of model parameters directly, leading to network congestion. In this invention, as a preferred approach, the consensus is split into a data diffusion layer and a metadata ordering layer. Nodes slice the model and extract Merkle root hashes, generating a Proof of Availability (CoA) after collecting sufficient signatures. In the ordering layer, nodes only asynchronously broadcast and topologically order the DAG vertices encapsulated with the CoA, achieving non-blocking ordering. The specific process is as follows:
[0068]
[0069] Algorithm 3 demonstrates the process of data decoupling diffusion and the generation of lightweight availability proofs. First, nodes slice and encode their large local gradient models using erasure coding (line 1), and construct Merkle trees based on these slices to extract a unique global identifier root hash (lines 2-3). In the subsequent data diffusion phase, nodes continuously send slices and their corresponding Merkle proofs to random nodes within the slices (lines 5-7). If a valid signature is received from a verification node, it is added to the local signature set (lines 8-10). When the number of collected signatures reaches the statutory fault tolerance threshold Q (line 12), the nodes use the BLS algorithm to merge and aggregate the signature set (line 13), and assemble it with the root hash and publisher information into a very small availability proof, which is then returned (lines 14-15), thereby decoupling physical data transmission from consensus ordering.
[0070] Cross-chip parallel sampling verification
[0071] To avoid the full download of parameters negating the bandwidth advantage of fragmentation during cross-fragment model synchronization, this invention, as a preferred approach, introduces a data availability sampling mechanism. Verification nodes only check the aggregate signature of the target fragment and randomly request a small number of data fragments and proof paths. As long as the sampled fragments align with the root hash, the full model can be proven to be truly usable with a high probability, reducing cross-fragment verification bandwidth consumption to logarithmic levels. The specific process is as follows:
[0072]
[0073] Algorithm 4 describes a parallel sampling mechanism for cross-shard verification of the legality of target model updates. Upon receiving the proof to be verified, the verification node first extracts the root hash and aggregate signature from the proof (lines 1-2), and uses the aggregate public key of the target shard to perform a preliminary verification of the signature's legality. If the verification fails, it is directly rejected (lines 3-5). Provided the signature is valid, the verification node does not need to download the full set of model parameters, but instead randomly requests a specified number of small data fragments and their Merkel proof paths from the target shard (line 6). Next, the node iterates through the sampled fragment set (line 7), verifying the cryptographic validity of each Merkel proof (line 10). If any fragment is found to be tampered with or cannot be aligned with the root hash, the verification is immediately rejected and an error is returned (line 11); if all sampled fragments successfully pass verification, it is highly probable that the underlying full model data is usable and secure (line 14).
[0074] Example
[0075] To verify the effectiveness of the core mechanism of the proposed ADES-FL scheme, simulation experiments were conducted to test the training performance of four datasets and models (0.1MB to 98MB) with increasing complexity, including MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100, with network communication latency (RTT) ranging from 50ms to 500ms, erasure coding configured as (16,12), and sampling size set to 8 to 48. The results are shown in Table 2.
[0076] The simulation experiment was conducted on the Ubuntu 20.04 operating system. A distributed federated learning environment was built based on Python 3.8 and PyTorch 1.9 frameworks, and Matplotlib was used for data visualization.
[0077] Table 2 Dataset Parameters
[0078]
[0079] To clearly verify the advantages of this method, the proposed scheme was applied to a distributed federated learning simulation platform and compared with other baseline schemes such as the traditional tightly coupled PBFT-FL algorithm and the DAG-BFL algorithm that partially implements asynchronous ordering.
[0080] System throughput and consensus latency analysis
[0081] To clearly verify the effectiveness of the asynchronous decoupling mechanism in handling large-volume parameters, the proposed scheme's performance in system throughput and consensus latency was compared with other schemes, as shown in Figures 1 and 2. Compared to PBFT-FL and DAG-BFL schemes, this method maintains higher system throughput under medium to heavy loads (44MB~98MB), and its throughput is almost unaffected by the increase in model size. Furthermore, the consensus latency of different schemes under the same high load pressure was verified. Under a 44MB load, the latency of this method remained stable at 1.2~2.5 seconds, and even under the extreme stress test of 98MB, it only slightly increased to 2.5~4.6 seconds. Therefore, this method has good concurrent ordering performance, effectively eliminating the blocking waiting time caused by the transmission of large-volume parameters, and significantly improving system consensus efficiency.
[0082] In the Internet of Vehicles (IoV), when processing large-volume autonomous driving models, vehicle nodes use erasure coding to slice local model parameters and spread them only among neighboring physical vehicle fleets or RSUs. Meanwhile, the RSUs, as representative nodes, asynchronously broadcast a very small volume of data availability proofs (CoA) on the directed acyclic graph (DAG) network. This completely decouples the transmission of large entity models from the ordering of consensus metadata, avoiding congestion and collapse of the vehicle's wireless channel.
[0083] Cross-chip communication overhead analysis
[0084] To further verify the performance of the cross-shard parallel sampling mechanism in reducing network communication burden, the total communication overhead of each scheme across the entire network in handling cross-shard interactions was statistically analyzed as the node size evolved, as shown in Figure 3. Under the extreme load of 98MB (CIFAR-100), the traditional architecture exhibits a strong positive correlation between communication overhead and model physical size, resulting in a large bandwidth consumption for a single cross-shard verification. Our proposed method successfully and strictly limits the single-node verification overhead to an upper limit of 5MB, achieving a 94.9% bandwidth reduction compared to the superior AS-BFL baseline scheme. This demonstrates that our method not only ensures the safe synchronization of model states but also achieves lower communication overhead under the same network resources, fully showcasing the effectiveness of its erasure coding sampling mechanism in alleviating the system's cross-shard bandwidth bottleneck.
[0085] Cross-regional verification nodes (such as RSUs in adjacent regions) do not need to download hundreds of megabytes of full autonomous driving model parameters from the target region. They only need to randomly request a very small number of physical data fragments and Merkel proofs to confirm the real availability of the underlying global model data with a high probability, which greatly saves the physical communication overhead of cross-regional interaction in vehicle-to-everything (V2X) communication.
[0086] Robustness and Byzantine fault tolerance assessment
[0087] To verify the fault tolerance baseline of the proposed system under extreme malicious environments, experiments were conducted to test the evolution of global model accuracy when different proportions of Byzantine nodes were injected into the network in a stepped manner. The results are shown in Figure 4. When malicious nodes cross the 20% high-risk threshold, the aggregation accuracy of the traditional scheme drops sharply. However, in heavy tasks such as CIFAR, even when facing malicious poisoning and data withholding attacks from Byzantine nodes with a proportion as high as 30%, the proposed method can still maintain a stable final convergence accuracy of over 88.5%, demonstrating the best performance. The wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding proposed in this invention not only achieves significant results in throughput, latency, and bandwidth reduction, but also greatly improves the security and fault tolerance of the system in the face of high-proportion malicious attack environments.
[0088] To address the problems of communication congestion, consensus blocking, and topology rigidity in large-scale distributed federated learning in wide area network (WAN) environments, this invention proposes a WAN federated learning algorithm based on asynchronous decoupling and elastic sharding. This method mitigates the impact of WAN latency fluctuations by introducing a virtual spatiotemporal coordinate system for elastic sharding. Furthermore, by constructing an asynchronous DAG pipeline that decouples data and consensus, it breaks through the scalability bottleneck of traditional synchronous Byzantine consensus. Combined with a data availability sampling protocol, it significantly reduces the communication overhead of cross-shard interactions. Experiments show that this method not only achieves significant results in reducing throughput, latency, and bandwidth, but also demonstrates excellent system robustness against high-proportion malicious attacks, making it highly valuable for practical engineering deployment.
[0089] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.
Claims
1. A wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding, characterized in that, include: Step 1: Measure the round-trip time of physical nodes based on the mapping algorithm and map it to virtual space coordinates. Divide all network nodes into multiple high-cohesion shards based on coordinate distance. Step 2: Train the local model based on all nodes in the network, generate local gradients, slice the gradients using erasure coding, construct a Merkle tree, spread the fragments within the slices, collect a quorum of verification signatures, and generate a lightweight data availability proof. Representative nodes from each shard asynchronously broadcast and topologically sort the DAG vertices that encapsulate the data availability proof, outputting a globally strict order. Step 3: Randomly request a small number of data fragments and Merkel proofs from the target shard, perform cross-shard sampling verification, merge the updates that pass the cross-shard verification into the global model, discard and mark malicious nodes if they cannot be executed; repeat steps 2 and 3 until the global model is output.
2. The wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding according to claim 1, characterized in that, Step one includes: Based on the virtual spatiotemporal coordinate system, each physical node in the entire network iteratively updates its own coordinates along the vector direction based on the error between the actual delay and the predicted delay; During the phase when the virtual coordinates are not converged, each node in the wide area network federated learning system randomly selects other nodes to probe its actual round-trip communication delay and calculates its Euclidean distance in the virtual space, using it as the basis for adaptively iteratively updating the virtual coordinates of the node. Once all node coordinates converge, the spatial coordinates are extracted, and the entire network of nodes is divided into K highly cohesive fragment sets based on the K-means clustering algorithm. For each sub-partition, find the node closest to its centroid as the representative node; Outputs shards and representative nodes used to accelerate consensus.
3. The wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding according to claim 2, characterized in that, Step two includes: Based on erasure coding technology, ordinary nodes slice and encode large-volume local gradient models. Construct a Merkle tree based on slice fragments and extract the global identifier root hash; Ordinary nodes continuously send fragments and Merkel proofs to random nodes within the shard. When a valid signature is received from a verification node, it is added to the local signature set. When the number of collected signatures reaches the statutory fault tolerance threshold Q, ordinary nodes use the BLS algorithm to merge and aggregate the signature set, assemble it with the root hash and publisher information into an availability proof, and return it to the representative node of this shard, thus completing the decoupling of physical data transmission and consensus ordering.
4. The wide area network federated learning algorithm based on asynchronous decoupling and elastic sharding according to claim 3, characterized in that, Step three includes: The availability proof is used as the proof to be verified. The root hash and aggregate signature are extracted from the proof. Based on the aggregate public key of the target shard, the legality of the signature is initially verified by the BLS signature verification algorithm. Under the condition that the signature is valid, the verification node only randomly requests a specified number of data fragments and their Merkel proof paths from the target shard. The verification node traverses the sampled fragment set and verifies the cryptographic validity of its Merkel proof one by one. If any fragment is tampered with or cannot be aligned with the root hash, the verification will be immediately rejected and an error will be returned. If all sampled fragments pass the verification successfully, it is highly probable that the underlying full model data is usable and secure.