Hangzhou Hikvision Digital Technology Co., Ltd.

Xiaohong Guan

Wei Wang

Bin Wang

Chao Li

Kezhang Lin

Li Wan

Haonan Yan

A systematic security protection method for a multimodal AI model is provided. The method comprises: in a security detection stage of the target multimodal AI model, according to a perceptual distance between detection data and original data determined by a trained distance-aware judgment model, obtaining detection data whose perceptual distance from the original data meets a requirement, and detecting the target multimodal AI model by using the detection data; in a case that the detection result in the detection stage is that the target multimodal AI model has a security problem, on one hand, performing security coding protection and tokenization filtering on the input of the model, on the other hand, performing multi-dimensional omnidirectional evaluation on the multimodal AI model.

Growing community of inventors

Hangzhou, China

Xiaohong Guan