Method for enhancing defense capability of neural network based on federated learning

Abstract

The invention provides a method for enhancing the defense capability of a neural network based on federated learning. The method comprises the following steps: S1, retaining data locally and preventing data privacy leakage by utilizing federated learning, cooperating with each party to carry out distributed model training, encrypting an intermediate result to protect data security, summarizing andfusing a multi-party model to obtain a federated model; and S2, establishing an adversarial sample, and quickly searching the adversarial sample by adopting an algorithm. According to the method, federated learning and a training process of a neural network model are combined, the problem that a data set cannot be circulated due to consideration of privacy protection and laws and regulations is solved, the trouble of data collection is saved, the training set of the neural network model is richer and more independent, and the defect that the neural network model is likely to be attacked by countermeasure samples due to the fact that the training set is incomplete is overcome.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence security, in particular to a method for enhancing neural network defense capabilities based on federated learning. Background technique [0002] Adversarial Examples (Adversarial Examples) refer to the addition of subtle disturbances that are difficult to distinguish artificially in the input samples, causing machine learning models (such as neural networks) to give a wrong output with high confidence. The existence of adversarial examples proves that existing machine learning models still have security problems, thus limiting the application and development of artificial intelligence (AI) in areas with high safety requirements (such as autonomous driving). The paper "Intriguing Properties of Neural Networks" (Christian Szegedy, Wojciech Zaremba, et al. Intriguing Properties of Neural Networks. In ICLR, 2014.) proposed the concept of adversarial samples and proved that adversarial ...

AI Technical Summary

Problems solved by technology

Therefore, the traditional data collection and use methods are no longer applicable, and big data-driven artificial intelligence is facing a data crisis. The method of simply expanding the data set and retraining the neural network model faces security risks such as data sharing and privacy leakage.

Images