Malicious clients simulation

[gautamjajoo]

No description provided.

[tremblerz]

I wonder if we should be creating a deeper layer of abstraction in the sys_config where each node gets assigned a corresponding algorithm configuration (the algorithm that it will run) and then we can create an algorithm and corresponding algorithm config file that malicious nodes run.

Obviously, to prevent cumbersome assignment of same algorithm to hundreds of nodes, this method of configuration will have to be layered on top with another parameter that automatically assigns the algorithms to each node. I have done something similar for GPU assignment already in get_device_ids https://github.com/aidecentralized/sonar/blob/main/src/configs/sys_config.py#L203

Advantages --

1. Each node can run a different algorithm so we can have heterogeneous mixes of different algorithms such as one malicious node and one free rider node etc.
2. The current implementation is a risky one because it relies on the fact that random function will work exactly the same way for everyone.
3. This abstraction layer/API will work much better when we actually run these things in a more decentralized manner where someone starts n nodes of some type and someone else starts m nodes of some other type.

Disadvantages --

1. Does not make sense to do this if we do not anticipate many use-cases with a mix of different kinds of nodes.
2. Our data partitioning system also relies on the random function working exactly the same way for all the nodes.

[tremblerz]

please only use the word nodes and not clients anywhere

[tremblerz]

num_users

[tremblerz]

Why are these two getting imported?

[tremblerz]

This should go in sys_config.py

[gautamjajoo]

I understand your point, and it seems a better solution. But as mentioned, if we do not plan to run with a mix of different kinds of nodes, we will lose time here.

I will start digging and check whether running with a mix is feasible or not

[gautamjajoo]

Now here is the structure:

1. There is sys_config which will assign a node whether it will be doing traditional_fl or malicious_traditional_fl. This is set as random. This is a map containing a node and its algo type.
2. Then inside the algo_config, each node if it's malicious_traditional_fl will be randomly assigned any of the malicious types, and if it is traditional_fl it will be working in its usual way.
3. The different malicious types are defined in malicious_config.py.
4. Also introduced hex uuid for storing the expt folder inside the expt_dump directory