Federated learning avoids centralizing data in a central server by distributing the model training process across devices, thus protecting privacy to some extent. However, existing research shows that model updates (e.g., gradients or weights) exchanged during federated learning may still indirectly leak sensitive information about the original data. Currently, single-key homomorphic encryption methods applied in federated learning cannot solve the problem of privacy leakage that may be caused by the collusion between the participant and the federated learning server, whereas existing privacy-preserving federated learning schemes based on multi-key homomorphic encryption in semi-honest environments have deficiencies and limitations in terms of security and application conditions. To this end, this paper proposes a privacy-preserving federated learning scheme based on multi-key fully homomorphic encryption to cope with the potential risk of privacy leakage in traditional federated learning. We designed a multi-key fully homomorphic encryption scheme, mMFHE, that encrypts by aggregating public keys and requires all participants to jointly participate in decryption sharing, thus ensuring data security and privacy. The proposed privacy-preserving federated learning scheme encrypts the model updates through multi-key fully homomorphic encryption, ensuring confidentiality under the CRS model and in a semi-honest environment. As a fully homomorphic encryption scheme, mMFHE supports homomorphic addition and homomorphic multiplication for more flexible applications. Our security analysis proves that the scheme can withstand collusive attacks by up to $N-1$ users and servers, where N is the total number of users. Performance analysis and experimental results show that our scheme reduces the complexity of the NAND gate, which reduces the computational load and improves the efficiency while ensuring the accuracy of the model.

The need of data owners for privacy protection has given rise to collaborative learning, and data-related issues heterogeneity faced by federated learning has further given rise to clustered federated learning; whereas the traditional privacy-preserving scheme of federated learning using homomorphic encryption alone fails to fulfill the privacy protection demands of clustered federated learning. To address these issues, this research provides an effective and safeguarded answer for sharing homomorphic encryption keys among clusters in clustered federated learning grounded in conditional representative broadcast re-encryption. This method constructs a key sharing mechanism. By combining the functions of the bilinear pairwise accumulator and specific conditional proxy broadcast re-ciphering, the mechanism can verify the integrity of homomorphic encryption keys stored on cloud servers. In addition, the solution enables key management centers to grant secure and controlled access to re-encrypted homomorphic encryption keys to third parties without disclosing the sensitive information contained therein. The scheme achieves this by implementing a sophisticated access tree-based mechanism that enables the cloud server to convert forwarded ciphertexts into completely new ciphertexts customized specifically for a given group of users. By effectively utilizing conditional restrictions, the scheme achieves fine-grained access control to protect the privacy of shared content. Finally, this paper showcases the scheme’s security against selective ciphertext attacks without relying on random prediction.

Edge drones continuously collect sensitive information such as telemetry data during missions, making it difficult to apply centralized model training directly due to privacy protection, security compliance, and regulatory constraints. Although federated learning (FL) can avoid sharing raw data, existing federated learning schemes based solely on homomorphic encryption (HE) still face security risks in drone scenarios, such as gradient inversion, member inference, and malicious update injection. To address this, we propose a secure and verifiable edge federated learning framework for parameter-efficient model adaptation in drone scenarios. The framework introduces homomorphic encryption for model updates on the device side to protect the privacy of updates before transmission and aggregation. Simultaneously, on the server side, decryption, aggregation, and verification are performed through a remotely authenticated Trusted Execution Environment (TEE), thereby limiting the server's access to plaintext updates and reducing the feasibility of gradient inversion and member inference attacks at the system level. Furthermore, an aggregation signature mechanism is introduced to batch verify the identity and update integrity of participating nodes, effectively preventing malicious or tampered updates from participating in aggregation, thus overcoming the shortcomings of existing HE-FL schemes in terms of poisoning resistance and verifiability. Experimental results show that, while ensuring safety and verifiability, the proposed method improves model accuracy by 3% compared to the comparative scheme, while maintaining better performance in terms of computation and communication overhead, thus verifying the practicality and deployability of the framework in resource-constrained UAV edge environments.

To address the security challenges of untrusted servers and the difficulty in verifying malicious node identities in UAV swarm federated learning when relying solely on homomorphic encryption (HE), this paper proposes a verifiable HE-based federated learning (VHEFL) framework that integrates trusted execution environments (TEEs) with aggregate signatures. The method proceeds through a series of well-defined steps. Initially, each UAV computes model gradients locally, encrypts them using additive HE, and attaches a lightweight identity-based digital signature. The encrypted model gradients and signatures are uploaded to the server. Subsequently, the ground server performs gradient aggregation, including both the encrypted gradients and signatures, in the regular execution environment (REE). After performing aggregation in the REE, the aggregated ciphertext and all the signatures are sent to the TEE. Inside the TEE, an efficient aggregate verification algorithm is executed to batch-verify the authenticity and integrity of all signatures simultaneously. This enables the TEE to securely verify the aggregated result and return the aggregated model parameters to the UAVs, where the encrypted model parameters are decrypted locally to obtain the updated parameters. By combining low-overhead aggregate signature verification with hardware isolation, VHEFL provides a high-performance and verifiable security solution that effectively addresses the challenges of data privacy, malicious update prevention, and computational overhead in UAV swarm federated learning.

With the rapid development of wireless communication and big data analysis technologies, the storage of massive amounts of data relies on third-party trusted storage, such as cloud storage. However, once data are stored on third-party servers, data owners lose physical control over their data, making it challenging to ensure data integrity and security. To address this issue, researchers have proposed integrity auditing mechanisms that allow for the auditing of data integrity on cloud servers without retrieving all the data. To further enhance the availability of data stored on cloud servers, multiple replicas of the original data are stored on the server. However, in existing multi-replica auditing schemes, there is a problem of server fraud, where the server does not actually store the corresponding data replicas. To tackle this issue, this paper presents a formal definition of authentic replicas along with a security model for the authentic-replica sampling mechanism. Based on time-lock puzzles, identity-based encryption (IBE) mechanisms, and succinct proof techniques, we design an authentic replica auditing mechanism. This mechanism ensures the authenticity of replicas and can resist outsourcing attacks and generation attacks. Additionally, our schemes replace the combination of random numbers and replica correspondence tables with Linear Feedback Shift Registers (LFSRs), optimizing the original client-side generation and uploading of replica parameters from being linearly related to the number of replicas to a constant level. Furthermore, our schemes allow for the public recovery of replica parameters, enabling any third party to verify the replicas through these parameters. As a result, the schemes achieve public verifiability and meet the efficiency requirements for authentic-replica sampling in multi-cloud environments. This makes our scheme more suitable for distributed storage environments. The experiments show that our scheme makes the time for generating copy parameters negligible while also greatly optimizing the time required for replica generation. As the amount of replica data increases, the time spent does not grow linearly. Due to the multi-party aggregation design, the verification time is also optimal. Compared to the latest schemes, the verification time is reduced by approximately 30%.