Cvpr2022 | backdoor attack based on frequency injection in medical image analysis

This work was jointly completed by JD Exploration Research Institute and Northwestern Polytechnic University , Has been CVPR2022 receive . In this paper, we propose a backdoor attack method based on frequency domain information injection （Frequency-Injection based Backdoor Attack,FIBA）. say concretely , We design a frequency domain trigger , By linearly combining the amplitude spectra of the two images , Inject the low-frequency information of the trigger image into the toxic image . because FIBA The semantics of contaminated image pixels are preserved , Therefore, we can attack the classification model , It can also attack the dense prediction model . We have carried out experiments on three benchmarks in the field of medical images ( For classification of skin lesions ISIC-2019 Data sets , For segmentation of renal tumors KiTS-19 Data sets , And for endoscopic artifact detection EAD-2019 Data sets ), To verify FIBA The effectiveness of attack medical image analysis model and its advantages in bypassing backdoor defense .

Thesis link ：https://arxiv.org/abs/2112.01148

Code link ： https://github.com/HazardFY/FIBA