Research On Privacy Security Of Federated Learning Based On Blockchain

The development of artificial intelligence,which is immensely data-dependent,has been hampered by growing concerns about data protection,privacy,and security,as well as the application of data protection laws and regulations.Scholars have proposed federated learning as a way to extract value from data while maintaining user privacy and security.Federated learning combines data from different sources using privacy-preserving strategies to create global models collectively.This strategy respects data contributors’ anonymity and encourages participants to collaborate to design models that generate more advantages.Data poisoning,model poisoning,and member inference assaults are still possible with federated learning.Because of blockchain’s quick development and inherent security and immutability,the two can be linked in an organic way,with blockchain technology serving as a trusted platform for federated learning participants.The use of federated chains for access management can bring parties with "trust difficulties" together to securely and trustfully train models for the greater good.With the following core effort,this paper supports federated learning based on blockchain technology to increase the privacy and security of federated learning.Unlike traditional federated learning frameworks,blockchain is being offered as a means of empowering federated learning.The usage of blockchain in federal learning provides a trustworthy authorisation method and identity management function for many participants,especially in a federated chain,joining parties who do not know and trust each other to establish a secure and reliable collaboration mechanism.PriChainFL,a federated learning framework for blockchain and IPFS,was conceived and built to introduce IPFS.The system coupling is reduced by shifting the model storing function to IPFS.PriChainFL achieves new features of storage records and model traceability per generation without affecting prediction accuracy,effectively reduces block costs,improves the security and robustness of the system,and improves the ability to resist member inference attacks and model poisoning attacks,according to the experimental results.Based on the low detection rate of data fingerprint security threats in PriChainFL,a K-CAT-based data fingerprint authentication model IAF is proposed,which combines the advantages of K-Means and Cat Boost to efficiently predict the current security status of the training side and timely exclude the threat training side.The data fingerprint authentication model is critical for enhancing the scalability of government learning systems.The simulation findings show that IAF can accurately detect security threats from abnormal data fingerprints.