| Federated Learning（FL）promotes data collaboration across institutions and regions by allowing numerous participants to jointly train machine learning models,addressing the issue of data silos present in traditional centralized model training.FL has the potential to drive the intelligent development of fields such as healthcare,finance,and transportation by enabling model training while keeping data local.However,FL also presents some security risks,such as privacy leakage of participant training datasets and illegal access to global models.This dissertation comprehensively considers potential security vulnerabil-ities in FL entities and training processes and proposes several key technologies based on blockchain and cryptography to enhance FL’s trustworthiness,ensuring the integrity and verifiability of local models and addressing issues of training dataset privacy leakage and global model privacy leakage.The main contributions of this dissertation are as follows:（1）Blockchain-based FL identity authentication technology.During the FL training process,untrusted participants and attackers in the network may lead to privacy leakage of legitimate participants and the inability of the global model to converge.Identity authen-tication mechanisms can verify local models’integrity and their sources’authenticity to alleviate this problem.However,existing identity authentication schemes have shortcom-ings,such as centralization,requiring large storage space,difficult management,and only being suitable for small-scale trusted networks,making them unsuitable for FL.There-fore,this dissertation proposes a decentralized and lightweight anonymous FL identity au-thentication scheme（DAFL）.By utilizing Directed Acyclic Graph（DAG）blockchain and dynamic accumulator,a decentralized and simplified storage FL identity authentication framework is designed.A lightweight digital signature algorithm supporting batch veri-fication is proposed for participant identity authentication using small coefficient testing technology.In addition,participants interact anonymously,enabling anonymous com-munication.Furthermore,when malicious behavior occurs,their real identities can be traced and recovered.Security analysis proves that the proposed authentication algorithm is non-forgeable,and DAFL meets various security requirements.Experimental results show that DAFL achieves lower authentication overhead compared to existing authen-tication schemes.Moreover,even in the presence of different proportions of malicious attackers,DAFL’s global model still maintains high accuracy.（2）Blockchain-based Security aggregation technology for Horizontal Federated Learn-ing（HFL）.Servers can infer the privacy data of participants from the local models up-loaded by participants,posing a risk of privacy leakage for participants.Additionally,par-ticipants may intentionally or unintentionally upload invalid local models,hindering FL model convergence.Existing HFL secure aggregation schemes suffer from various issues,such as compromised model accuracy,significant security computing and communication overhead,and scalability issues due to hardware dependence.Therefore,this dissertation proposes a decentralized,secure,and verifiable HFL secure aggregation scheme（LPBFL）.Firstly,a blockchain-based hierarchical framework is proposed,introducing edge aggre-gation nodes between task publishers and participants to avoid performing global model update operations on local model ciphertext,greatly reducing privacy protection commu-nication overhead.Then,a digital signature algorithm is designed to resist man-in-the-middle attacks during the HFL training process.Finally,a participant reputation selection mechanism is designed using subjective logic models to make HFL more efficient.The-oretical analysis and experimental results demonstrate that LPBFL achieves lower com-munication overhead and higher global model accuracy compared to related HFL secure aggregation schemes while satisfying various security requirements and resisting tradi-tional attacks.（3）Blockchain-based privacy protection technology for Vertical Federated Learn-ing（VFL）.VFL enables multiple participants to train using overlapping user feature data jointly,but this process may lead to privacy leakage of participant features or labels.Exist-ing VFL privacy protection schemes have drawbacks,such as applying only to two partic-ipants,requiring a third-party coordinator,compromising model accuracy,and having ex-cessive computation or communication overhead.Therefore,this dissertation proposes a decentralized blockchain-based multi-party VFL privacy protection scheme（Secure VFL）.A verifiable2（4）-sharing protocol supporting four-party VFL joint training is designed us-ing Replicated Secret Sharing（RSS）.The protocol involves only addition operations and exhibits robustness.Additionally,a Proof of Feature Sharing（Po FS）consensus algorithm is proposed,where the participant who first achieves feature reconstruction is responsible for generating blocks,ensuring the credibility of VFL training while improving blockchain system efficiency.Security analysis proves that Secure VFL has good security and relia-bility.Experimental results show that Secure VFL has significant advantages in training efficiency and model performance compared to existing four-party VFL privacy protec-tion schemes.（4）Blockchain-based security sharing technology for FL.After training the global model（i.e.,knowledge）,deploying knowledge to many other devices may result in knowl-edge leakage.Existing sharing schemes may have issues such as the leakage of access policy privacy and users with revoked attributes still being able to decrypt information.Therefore,this dissertation proposes a privacy-preserving FL knowledge sharing scheme（CBKFL）.A verifiable knowledge sharing architecture is built using blockchain,and by optimizing the committee member selection method of the Delegated Proof of Stake（DPOS）consensus algorithm,issues such as Sybil attacks and bribery attacks are ad-dressed.Furthermore,a PHR-CP-ABE algorithm that supports fine-grained access control is designed,which has the functions of access policy hiding,attribute privacy protection and attribute revocation.Security analysis proves the Chosen Ciphertext Attack（CCA）security of CBKFL.Experimental results show that CBKFL has lower privacy protection overhead compared to existing sharing schemes. |
