Missing validation results in undefined behavior in `QuantizedConv2D`

Impact

The implementation of tf.raw_ops.QuantizedConv2D does not fully validate the input arguments:

import tensorflow as tf

input = tf.constant(1, shape=[1, 2, 3, 3], dtype=tf.quint8)
filter = tf.constant(1, shape=[1, 2, 3, 3], dtype=tf.quint8)

# bad args
min_input = tf.constant([], shape=[0], dtype=tf.float32)
max_input = tf.constant(0, shape=[], dtype=tf.float32)
min_filter = tf.constant(0, shape=[], dtype=tf.float32)
max_filter = tf.constant(0, shape=[], dtype=tf.float32)

tf.raw_ops.QuantizedConv2D(
  input=input,
  filter=filter,
  min_input=min_input,
  max_input=max_input,
  min_filter=min_filter,
  max_filter=max_filter, 
  strides=[1, 1, 1, 1],
  padding="SAME")

In this case, references get bound to nullptr for each argument that is empty (in the example, all arguments in the bad args section).

Patches

We have patched the issue in GitHub commit 0f0b080ecde4d3dfec158d6f60da34d5e31693c4.

The fix will be included in TensorFlow 2.9.0. We will also cherrypick this commit on TensorFlow 2.8.1, TensorFlow 2.7.2, and TensorFlow 2.6.4, as these are also affected and still in supported range.

For more information

Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.

Attribution

This vulnerability has been reported by Neophytos Christou from Secure Systems Lab at Brown University.