Accountable and Verifiable Secure Aggregation for Federated Learning in IoT Networks
In the Internet of things (IoT) networks, largescale IoT devices are connected to the Internet to collect users' data. As a distributed machine learning paradigm, federated learning (FL) collaboratively trains the global model by utilizing large-scale distributed devices, while protecting the privacy of the local data sets of each participant. Federated learning with secure aggregation employs an aggregation server (aggregator) to compute a multiparty sum of model parameter updates of each participants in a secure manner and further realizes the updates. However, existing schemes are usually based on semi-honest assumptions, which make them vulnerable to malicious clients. In addition, they address the random client dropouts problem by increasing the data size, which brings a large communication overhead. To solve these issues, we propose an accountable and verifiable secure aggregation for federated learning framework. Specifically, we employ an SMC protocol based on homomorphic proxy re-authenticators and homomorphic proxy re-encryption to execute secure aggregation, while integrating the blockchain to realize the function of penalty for malicious behavior. Our framework can guarantee the verifiability of data provenance and is accountable for malicious clients. To demonstrate the usability of our framework, we evaluate the specific cryptography schemes and develop a blockchain-based prototype system by using solidity language to test the performance of the framework.
Blockchain, Cryptography, Learning systems, Network security
X. Yang, Y. Zhao, Q. Chen, Y. Yu, X. Du and M. Guizani, "Accountable and Verifiable Secure Aggregation for Federated Learning in IoT Networks," in IEEE Network, vol. 36, no. 5, pp. 173-179, September/October 2022, doi: 10.1109/MNET.001.2200214.
IR Deposit conditions:
OA version (pathway a) Accepted version
When accepted for publication, set statement to accompany deposit (see policy)
Must link to publisher version with DOI
Publisher copyright and source must be acknowledged