Federated gradient consensus framework for secure and trust-adaptive collaborative model training across distributed nodes

Abstract

Federated gradient consensus framework for secure and trust-adaptive collaborative model training across distributed nodes, including: multiple client nodes that perform local training and gradient computation; a federated gradient consensus layer that validates gradient directions and forms a distributed consensus; a confidence rating unit that assigns dynamic confidence weights to the nodes; and a secure gradient aggregator that uses cryptographic or multi-party computational methods for privacy-compliant aggregation.

Description

Technical field of the utility model

[0001] The utility model relates to distributed machine learning systems, specifically a federated gradient consensus mechanism that enables secure, privacy-compliant, and trust-adaptive collaborative model training without centralized data exchange. The system integrates cryptographic aggregation, adaptive consensus, and reputation-based weighting to enhance model integrity in federated networks. Background of the utility model

[0002] Existing federated learning frameworks allow multiple participants to train shared models without exchanging raw data. However, non-IID datasets, unreliable participants, and malicious updates lead to unstable training, gradient drift, and data breaches. Centralized aggregation servers represent a single point of failure and are vulnerable to manipulation.

[0003] Therefore, there is a need for a secure, decentralized, and consensus-based gradient aggregation system that maintains the stability and fairness of the model while ensuring data privacy. Summary of the utility model

[0004] The following presentation offers a simplified summary of the disclosure to provide the reader with a basic understanding. This summary does not constitute a comprehensive overview of the disclosure and does not identify the essential or critical elements of the utility model, nor does it limit the scope of its protection. Its sole purpose is to present some of the concepts disclosed herein in simplified form as an introduction to the more detailed description presented later.

[0005] The present utility model introduces a Federated Gradient Consensus (FGC) framework that includes the following: a Federated Gradient Consensus Layer (FGCL) that verifies and adjusts local gradient directions through distributed consensus; a Trust Evaluation Unit (TEE) that assigns dynamic weights to each client based on its reliability and past performance; a Secure Gradient Aggregator (SGA) that uses homomorphic encryption, secure multi-party computation, or differential privacy; optionally, a blockchain-based ledger that ensures tamper-proof tracking of the consensus.

[0006] The present utility model enables secure, scalable and data protection-compliant collaborative model training between distributed units. Detailed description of the utility model

[0007] It should be understood that the present disclosure is not limited in its application to the design details set forth in the following description or to the arrangement of components, or to the embodiments shown in the drawings. The disclosure is suitable for other embodiments and can be implemented or carried out in various ways. Furthermore, it should be understood that the language and terminology used herein serve only for descriptive purposes and are not to be considered limiting. 1. System Overview: This utility model discloses a Federated Gradient Consensus (FGC) framework designed to enable secure, collaborative, and decentralized model training among multiple participants or organizations without requiring direct exchange of raw data. The framework combines principles of federated learning with cryptographic security, trust assessment, and consensus optimization. Each participating node contributes local model updates computed from private datasets. These are securely aggregated and validated through a distributed consensus process. This approach ensures data privacy, resilience against malicious clients, and stable model convergence in heterogeneous environments. 2. Architectural Components: The FGC framework consists of three main components: client nodes, a Federated Gradient Consensus Layer (FGCL), and an optional blockchain-based audit layer. Each client node independently trains a local model on its own dataset and computes gradients that represent the local learning direction. These gradients are encrypted and transmitted to the FGCL, where they undergo consensus validation and trust-based weighting. The optional blockchain layer serves as a transparent ledger that records consensus events, trust weights, and model updates for later review and verification. 3. Federated Gradient Consensus Mechanism: The federated gradient consensus layer establishes agreement between nodes regarding the direction and quality of gradient updates. Each client node transmits its gradient g. ito the FGCL. The system calculates a temporary consensus direction g. c , which represents the average or median of gradients from trusted nodes. A similarity metric, such as cosine similarity, is used to compare individual gradients to the consensus direction. Gradients falling below a predefined similarity threshold θ are flagged as inconsistent or potentially malicious and excluded from aggregation. The final aggregated gradient g agg is then calculated using confidence-weighted averaging to ensure robust convergence. 4. Trust Evaluation Engine: The Trust Evaluation Engine (TEE) assigns each node a dynamic trust weight α. ito, which is based on its past reliability and its contribution to model improvement. Three metrics are used: (a) the gradient consistency index (GCI), which measures directional similarity over multiple training rounds, (b) the contribution fidelity (CF), which assesses how much updating a node improves validation performance, and (c) the reputation score (RS), which describes the long-term reliability of the node. The combined confidence weight α i = f (GCI i , CF i , RS i ) ensures that nodes with unreliable or irregular behavior have less influence on the global aggregation. 5. Secure Gradient Aggregator: The Secure Gradient Aggregator (SGA) ensures privacy-compliant model aggregation using cryptographic methods. Each node encrypts its gradient using homomorphic encryption, allowing the aggregator to calculate the sum of the encrypted values ​​without decrypting them. Alternatively, Secure Multi-Party Computing (SMC) can be used, where each gradient is decomposed into random parts and distributed among the participants. Additionally, differential privacy can be applied by adding controlled random noise to the gradient updates. These mechanisms protect the confidentiality of local data and ensure compliance with data protection regulations such as the GDPR. 6. Blockchain Audit Layer (Optional Implementation): In certain implementations, a blockchain audit layer is integrated into the framework to ensure traceability and accountability. Each training round generates a block containing hashed identifiers of the aggregated gradients, the trust weights of the participating nodes, timestamps, and consensus signatures. This immutable ledger ensures transparent verification of model development, supports regulatory compliance, and prevents manipulation or denial in multi-institutional collaborations. 7. Adaptive Optimization and Convergence Control: The global model is adaptively updated based on the variance of the consensus gradients. If the consensus variance exceeds a threshold δ, the learning rate η is reduced, and stricter confidence weighting is applied. This dynamic adjustment ensures stable convergence, even under non-IID or adversarial training conditions. The framework supports asynchronous update schedules and gradient compression to optimize computational effort and bandwidth. 8. Implementation Example: In an exemplary implementation, the FGC framework is used in a consortium of hospitals that jointly train a diagnostic model. Each hospital maintains its own patient dataset and trains a local model instance. Gradients are securely exchanged using homomorphic encryption and consensus validation. The system prevents data leaks, protects patient confidentiality, and achieves accurate predictive performance through trust-based adaptive learning. 9. Advantages of the utility model The proposed framework offers several advantages: - Increased trust and security in federated learning, - Protection against data leaks and manipulation attacks, - Improved convergence with heterogeneous datasets, - Transparent and verifiable training records, - Scalability and low communication overhead for large-scale networks.

[0008] It must be acknowledged that the subject matter described above can also be embodied in other specific embodiments without deviating from the scope of protection or the essential features of the disclosure. Therefore, it should be understood that the subject matter is not intended to be limited by the preceding explanatory details, but rather is defined by the accompanying claims.

Claims

[1] Federated gradient consensus framework for safe and trust-adaptive collaborative model training across distributed nodes, comprising: multiple client nodes that perform local training and gradient computation; a federated gradient consensus layer that validates gradient directions and forms a distributed consensus; a confidence rating unit that assigns dynamic confidence weights to the nodes; and a secure gradient aggregator that uses cryptographic or multi-party computational methods for privacy-compliant aggregation. [2] Frame according to claim 1, characterized by that the consensus layer uses a Byzantine fault-tolerant or weighted median consensus algorithm. [3] Frame according to claim 1, characterized by, that the trust rating unit calculates the trust weights using gradient consistency, contribution fidelity, and reputation metrics. [4] Frame according to claim 1, characterized by that the secure gradient aggregator uses homomorphic encryption, secure multi-party computation, or differential privacy. [5] Frame according to claim 1, characterized by that it includes a blockchain-based audit ledger that records consensus events and aggregated updates. [6] Frame according to claim 1, characterized by , that global model updates are adaptively optimized based on the consensus variance to improve convergence stability.