Fix incorrect buffer index failure

[ilya-kupchenko]

Description:
bug=fixes #3075 #3076 #3077

Problem
MicroAllocator does not validate the buffer indices. If a model provides an out-of-bounds buffer index for an input, output, or intermediate tensor, it results in segmentation fault.

Solution
Add validation inside MicroAllocator to check that the buffer index for every tensor. If an invalid index is found, InitializeTfLiteTensorFromFlatbuffer and InitializeTfLiteEvalTensorFromFlatbuffer will now return kTfLiteError.

Files Modified
tensorflow/lite/micro/micro_interpreter.cc
tensorflow/lite/micro/micro_interpreter_test.cc

Testing
Added new Bazel tests that use models with invalid buffer-to-tensor mappings.

[ilya-kupchenko]

Hi, this is one of my first PRs. Sorry if anything is wrong (I tried my best). I added the validators for subgraph and tensor indices in MicroAllocator::PopulateTfLiteTensorFromFlatbuffer. I am not sure if this is needed and I am not sure about this:

  if (tensor_index < 0 ||
      static_cast<size_t>(tensor_index) >= subgraph->tensors()->size()) {
    MicroPrintf(
        "Invalid tensor index %d for subgraph %d, subgraph has only %d tensors"

I have heard that "-1" is used sometimes to indicate optional tensors or lacking dimensions, not sure if this should be caught as well.

Contributing on TFLM project was my dream, sorry if I did anything wrong.

[gonnet]

Thank you for fixing this!