Research On Privacy Protection Scheme Of Federated Learning Based On Functional Encryption And Edge Computing

Federated Learning has been widely researched by companies that have large amounts of user behavior data.In this approach,each company trains a local machine learning model by inputting user data.The model information from each company is then aggregated by a role called the aggregator to create a global model,which is distributed to each company for further training.However,concerns have been raised about privacy during the training and prediction stages,such as model inversion attacks and membership inference attacks.Additionally,companies are concerned that private model parameters may be leaked during the aggregation phase,leading to the aforementioned attacks and resulting in the leakage of a large amount of user information.To address these issues,various security and privacy protection technologies such as differential privacy and secure aggregation have been proposed.To address the privacy concerns mentioned above,people have proposed the use of Secure Multi-Party Computation(SMC)or Homomorphic Encryption(HE)in privacypreserving deep learning and federated learning schemes.However,these solutions suffer from efficiency issues,leading some to suggest the use of Functional Encryption(FE)in machine learning and federated learning scenarios.For existing local training schemes based on FE,the decryption phase accounts for the majority of the training time.Furthermore,current federated learning parameter aggregation schemes suffer from functional limitations,such as the lack of traitor tracing revocation,decentralization,and user dynamic exit functionality.In response to the aforementioned challenges and current situation,this paper proposes a privacy-preserving federated learning scheme that includes two stages: secure local training and parameter secure aggregation.This dissertation proposes a secure local training scheme that combines functional encryption and edge computing.Additionally,this dissertation introduces a new functional encryption primitive that supports traitor tracing revocation,decentralization,and user dynamic exit functionality,which are applied to our proposed secure aggregation scheme.The specific work is as follows:1.This dissertation proposes a functional encryption scheme based on edge computing and functional hiding,along with a local training scheme for functional encryption primitives.These schemes are used to protect user data and the model parameters of the training party.The decryption task can be assigned to edge nodes,allowing for parallel execution of decryption and training on user data,thereby improving system efficiency.2.This dissertation proposes a dynamic decentralized function encryption scheme that is threshold traceable and revocable,using traceable secret sharing,Non-Interaction key exchange,and authenticated encryption.The proposed scheme is proven to be selective indistinguishability(sel-IND)under the Decisional Diffie-Hellman(DDH) assumption.3.This dissertation proposes a secure aggregation scheme for federated learning based on threshold,traceable,revocable,and dynamically decentralized functional en-cryption.The scheme enables users to generate keys in a decentralized manner during the aggregation phase,allowing them to exit the system at any time.Ad-ditionally,when traitors exist,they can be traced and their aggregation privileges revoked.Compared to previous schemes,This dissertation has an advantage in functionality.