search

CoderWZW / ARLib

An open-source framework for conducting data poisoning attacks on recommendation systems, designed to assist researchers and practitioners.
86 stars 11 forks source link

about fedrecAttack and Legup implementation #17

Open qqt-code opened 2 months ago

qqt-code commented 2 months ago

I debugged the fedrecattack code and referred to the original implementation, and it seems that there are some issues with this part of the implementation.about 80 row in fedrecattack.py,I believe this section of the code should be modified below： image and in Legup.py，i read the paper and i think the loss should L_GD - L_RS?,I'm sorry to bother you, but I wanted to ask if I misunderstood something.

qqt-code commented 2 months ago

I am sorry to bother you again, Why am I only getting poor experimental results? Could it be that my parameter settings are incorrect?I have setted maliciousUserSize from 0.01 -> 0.1,and adjust attackTargetChooseWay，but still got poor results，if there are any other hyperparameter to adjust? thank u very much.

M1234543 commented 2 months ago

Thank you very much. Based on your suggestions, we found that there is indeed an issue in FedRecAttack. Our intention was to find the row and column indices of the non-zero values from the matrix uiAdj2. Your code correctly achieves this functionality, but we opted to use the .nonzero() method instead of the indices attribute. Regarding LegUp.py, there is indeed a problem with its implementation. We noticed that in the original paper, only L_gen was optimized at the end of step k2, and we have made corresponding changes in the code. Concerning the issues you mentioned with the experimental results, could you provide the parameter settings of the methods you ran so that we can better test the code?