Image recognition back door defense method and device based on swarm intelligence, and medium

Abstract

The invention discloses an image recognition backdoor defense method and device based on swarm intelligence, and a medium. The method comprises the following steps: selecting neural network models with dense connection as seed neural network models based on the incidence relation of a plurality of neural network image recognition models in a distributed cluster and according to the ranking of connection density; aiming at every two seed neural network models, performing distillation operation on the two seed neural network models by utilizing attention distillation representation, and extracting an attention activation graph Al which is invalid to the back door attack of the seed neural network models; and measuring the Euclidean distance and the cosine distance of activation attention maps Al of the two seed neural network models by designing a robust distillation loss function L, allowing the seed neural network models to calculate gradient values based on the L and to carry out back propagation, and updating the parameters of the seed neural network models. The performance and efficiency of an image recognition back door defense algorithm based on swarm intelligent under the image recognition back door defense task reach a current optimal level.

Description

technical field [0001] The invention relates to the field of artificial intelligence, in particular to an image recognition backdoor defense method, device, product and storage medium based on swarm intelligence. Background technique [0002] The application of neural network image recognition in automatic driving and medical diagnosis has been developed rapidly. However, since most neural network image recognition is optimized in a performance-driven manner, it does not pay much attention to whether the neural network will be attacked by malicious parties. tamper. There are growing concerns about the safety of neural networks in image recognition applications. Due to the opacity of neural networks, adversarial attacks and backdoor attacks have become the main factors threatening the security of neural networks. Adversarial attacks mislead the model to predict wrongly by crafting adversarial examples, however, crafting strong adversarial training examples is computationall...

AI Technical Summary

Problems solved by technology

[0005] (1) The existing image recognition backdoor defense methods have not explored and studied the defense efficiency of backdoor attacks, making it difficult to effectively defend in a distributed cluster environment in a real environment

[0006] (2) In the traditional image recognition neural network, the image recognition model that has not been implanted into the backdoor trigger can be trained one by one by using the image data that has not been implanted into the backdoor trigger to erase the trigger of the backdoor attack, which greatly limits the eradication removal efficiency

Images