fuchong-cyber / label-inference-attacks Goto Github PK

In your paper, you reported that for CIFAR-10 dataset, the main task accuracy is 0.8280 while the passieve model completion attack performance is 0.8024 on the training dataset and 0.6299 on the testing dataset with 40 auxiliary labeled data in total.

However, when re-running your code on my device (A100) without any modification of the default hyper-parameter settings, I got 0.8015 for main task performance, which is similar to your reported result, but 0.8721 on training dataset and 0.6984 on testing dataset for attack accuracy, which is apparently too high compared to your reported results.

Could you help me explain why this could happen?

您好，打算复现原论文中的模型补全算法，想问一下run_training.bat与run_model_completion.bat中的三行指令分别代表什么呢？还有跑python vfl_framework.py --use-mal-optim True --use-mal-optim-all False --use-mal-optim-top False -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --epochs 100 --half 16 该命令时，loss值始终为NAN，请问这是正常现象吗

Hey!
I am new to VFL. I want to run this code but I didnt understand the instructions given in read.md file.
Kindly guide me with all the steps to run this code for only single attack(active attack) on windows system.
Thanks in advance..

I want to comfirm my understanding in your code. In ReadME indicates that 'run_model_completion.bat' will run the passive and active label inference attacks. The train() function in model_completion.py is to train a top model for passive label inference attack right? And the validation code to check accuracy on complete training set and testing set also showing the metrics score of passive label inference attack?
Thanks

看上去只有第一个step需要append，后面的step append也没用，还是用的第一次append进去的list

Hello, I am trying to reproduce the results of the defense strategy against label inference attacks on 4090. However, I find that the training speed is too slow. I wonder is there something wrong?

This is the running script:
python vfl_framework.py --use-mal-optim True --use-mal-optim-all False --use-mal-optim-top False -d CIFAR10 --path-dataset /Datasets/CIFAR10 --k 5 --epochs 100 --half 16 --lap-noise True --noise-scale 1e-3

I am a beginner in deep learning and have read the code carefully. But there are still some questions I would like to ask.
In the backpropagation phase of the vertical federated framework, the bottom model needs to calculate a loss. And the loss is

The note states

read grad of : input of top model(also output of bottom models), which will be used as bottom model's target

I didn't understand. Why the bottom model's loss is calculated in this way?

I am trying to run vfl_framework.py based on the dataset Yahoo Answer. I have downloaded the Bert model 'bert-base-uncased' and saved it in models/transformers/. However, something is still wrong.