A trusted consensus fusion scheme for decentralized collaborated learning in massive IoT domain

Highlights

• Proposed a trusted consensus fusion scheme for federated learning in IoT systems.

• The scheme combines trust evaluation system with X-BFT to build a trust layer.

• The trust layer helps to evaluate trust level and maintain consensus fusion stable.

• A trust reward & punishment method is designed for trust incentive consensus.

• Developed security models for Trusted X-BFT and performs security analysis.

Abstract

In a massive IoT systems, large amount of data are collected and stored in clouds, edge devices, and terminals, but the data are mostly isolated. For many new demands of various intelligent applications, self-organized collaborated learning on those data to achieve group decisions has been a new trend. However, in order to reach the goal of group decisions, trust problems on data fusion and model fusion should be solved since the participants may not be trusted. We propose a consistent and trust fusion method with the consortium chain to reach a consensus, and complete the self-organized trusted decentralized collaborated learning. In each consensus process, consensus candidates check others’ trust levels to ensure that they tends to fuse consensus with users with high trust, where the trust levels are evaluated by scores according to their historical behaviors in the past consensus process and stored in the public ledger of blockchain. A trust rewards and punishments method is designed to realize trust incentive consensus, the candidates with higher trust levels have more rights and reputation in the consensus. Simulation results and security analysis show that the method can effectively defend malicious users and data, improve the trust perception performance of the whole federated learning network, and make the federated learning more trusted and stable.

Introduction

In the scene of massive IoT systems, there are many demands of data fusion and cooperation between devices[1], [2]. For example: multiple devices with local data converge to the global artificial intelligence model; in intelligent transportation, multiple vehicles help to feedback traffic information; multi-agent joint decision-making. Currently, making full use of terminal devices, edge devices and cloud data centers to achieve collaborative decision-making has been a new trend. However, most of the current distributed learning frameworks require parameter servers to aggregate local model updates. Unfortunately, these server based frameworks suffer from a single point of failure. In addition, for example, the federal learning participants can obtain the same final model regardless of their specific motivations and contribution, which leads to the blindness and unfairness. In other words, in the current distributed learning architecture, all participants can receive the same global model at the end of collaborative model training. Actually, the participants in collaborated leaning needs self-organized for their specific common requirements. Thus, the final models would benefit them more.

In order to realize self-organized collaborated learning for massive IoT systems, the basis of cooperation is the trustworthiness between partners, that is, the information provided by the other party and the decisions made are trusted and reliable. There are two ways to build trust. The former is based on cryptographic techniques to provide protection against outside attacks by managing authentication and access control processes and by safeguarding privacy[3]. Generally, a central management organization is set up in a domain to manage or verify the identity in a distributed form. If the identity is verified to be reliable, the allowed behavior is trustworthy. However, how to conduct mutual authentication of identities is a difficulty. The other one is based on distributed reputation systems that estimate the degree of trust of a node by exploiting the opinion of others (reputation), usually arranged in a single synthetic measure as, for instance, in[4]. However, both the reliability of the witness and the relevance of the opinions are problems to solve.

Blockchain ensures the authenticity of recorded data, with its unforgeability, and helps to confirm their reliability with recorded data, and confirm their traceability and consensus mechanism in totally decentralized way, so it can better build trust in cross domain/open network. The records in these blocks are tamper proof[5], [6], [7]. While, it provide a measure to store model parameters of model fusion[8], [9], [10], [11]. Consensus protocol is the core of blockchain. The X-BFT[12] consensus protocols are usually used in the consortium chain-based federated learning systems, it is a type of consensus protocols matching federated learning. Byzantine fault tolerance (BFT) is a fault tolerance technology in the field of distributed computing. In recent years, various BFT consensus protocols (X-BFT) have been applied to the consortium chain platform. The BFT protocol is based on the message passing between nodes to reach a consensus on the protocol in the network. Once a consensus is reached, a certain result is formed. At present, the BFT consensus protocols have been applied in financial systems, supply chain systems, and so on[13], [14], [15].

However, the X-BFT protocols themselves are facing trust and security challenges[16]. It lacks security analysis of blockchain-enabled systems, including the security of the consensus algorithm, incentive mechanism, smart contract, and other key links in the blockchain. Besides, the consensus agreement is processed without referencing nodes’ historical behavior, which means that even some nodes have committed malicious or selfish behavior, they can also be trusted in the next consensus procedure[17].

However, in order to realize blockchain-enabled trusted collaborated learning, trust evaluation is the key problem. The X-BFT protocols did not include trust management function, which is quite important for consensus fusion in collaborated learning. We propose a Trust X-BFT (TX-BFT) protocol, which combines the trust evaluation system with the X-BFT through building a trust layer to evaluate trust level of consensus candidates and maintain consensus fusion stable and reliable. In order to evaluate our scheme, we built security models for TX-BFT, performed security analysis and a series of simulations. With trust evaluation and validation, consensus in collaborated learning can be improved in trust domain. As a result, the collaborated learning systems can be more trusted and reliable.