Research on Block Chain Defense against Malicious Attack in Federated Learning
Ming Wu, Lu Ge Hao, Fu Li Yu, Peng Mao · 2021
Federated learning enables participants to be capable of collaboratively building powerful machine learning models and exploiting privacy protection mechanisms to protect data privacy. However, the federal-learning security mechanism remains not perfect. Using malicious training samples is considered to be vulnerable to attacks against machine learning models. Lastly, the substitution technique of the model is adopted to eliminate the learning effect of the final model. Though numerous security protocols have been formulated to defend against and prevent malicious attacks, most are active defenses against malicious attacks, instead of passive ones (e.g., distillation-based defense and regularization-based adversarial training). In the present study, a block chain-based passive defense model is proposed to protect the model. To be specific, when a participant updates its local model, model participation, user fingerprints and other key information will be stored on the chain. Malicious attacks can be therefore traced and detected. To ensure data integrity and confidentiality, model data is encrypted and updated via the blockchain. Record rewards for active participation in federal learning and punish violators. Accordingly, the dynamic protection mechanism of federated learning is achieved.