anupamme / fair_flearn Goto Github PK

There is another implementation of SVM in tensorflow as pointed out in this SO link.

However, the existing implementation in flearn/models/vehicle/svm.py does not use this implementation.

It seems that the implementation in SO link is more native so we want to experiment with this and understand.

An adversary needs to make 3 choices:

1. How many nodes to contribute
2. What fraction of data to poison
3. What is the poison to add

We will focus on first 2 questions in this issue. We need to make changes in the code so that we can control the values to these questions e.g.

An adversary decides to contribute 10 nodes and 20% poisoned data in each of them - this means create adversarial data for this requirement.

How to create data for such a requirement. There is a companion issue [1].

Make changes in the existing code base to accommodate for adversarial clients - I have added placeholders there.

1. jaintj95/Federated_AoT#3