Bit Flips in the weights

[abhishektyaagi]

Hi,
As far as I understand, TensorFI is capable of injecting faults at the level of nodes but not in the nodes rather the output of one node acts as the input to the other nodes.

If we want to study the effects of bit flips in the weights of a model, how can we achieve it with TensorFI?

Is there a way to inject the bit flips inside the 'weight node'?

[nniranjhana]

Hi @abhishek-t-naive thanks for your question - we consider injection in the weights in our latest version of the tool: TensorFI2 which works with TensorFlow 2 and Python 3.

[abhishektyaagi]

That's great to hear!

I am trying to work with a network implemented with TensorFlow 1 and Python 2.7. Do you suggest any method by which injection in weights could be done with TensorFlow 1 or TensorFlow 2 back-compatible?

[nniranjhana]

I see.. right now, I don't think we have a solution for your use case but I see this as a possible enhancement for TensorFI 1, so adding the label.