The issue of inconsistent linearity results

[sincere0909]

Thank you very much for providing the open-source project. However, I encountered some confusion while using your code and would appreciate your help.
First, I ran the command “CUDA_VISIBLE_DEVICES=0 python model_train.py --dataset cifar10 --network resnet18 --phase train” to train a clean model, and the accuracy was normal. Then, I ran the command “CUDA_VISIBLE_DEVICES=0 python eval_linearity.py --attack clean --dataset cifar10 --network resnet18” to evaluate the linearity of the clean model. However, the result shows a linearity score of 0.9911737153566863, while the result in Table 4 of the original paper is 0.47. This discrepancy is quite significant, and I am unsure about the cause.
I look forward to your reply and greatly appreciate you.

[KaiyuanZh]

Hi @sincere0909 , thank you for your interest in our work, you can play with the _k and the linear_inputs parameters. The linearity scores of a clean model can fluctuate more and have lower values, and backdoored models yield stronger linearity.