Memory-based Fault Injection into weights and biases

[vanessaip]

Week 1-4 (sep 23 - oct 18): inject bit flip fault into a random weight of every fully connected layer in a simple model
Week 5-6 (oct 21 - nov 1): inject bit flip into biases of every fully connected layer and be able to select exactly which weight in a tensor to modify (for testing purposes)
Week 6-7 (nov 4- nov 15): Can inject faults into select layers
Week 8-9 (nov 18 - nov 29) : perform evaluation of SDC rate and performance overhead
Week 10-12 (dec 2 - dec 13): explore injecting faults into weights of other model architectures like CNNs and Transformers
Week 13 (dec 16 - 20): buffer time for further evaluation or implementation

week 5 update:

• injecting into first source register of fmul instruction any weight in matmul layers && verify that weights were modified

[karthikp-ubc]

Just following up on this. It'd be great if you're able to integrate this feature into the main LLTFI branch if it is done. Thanks.