CVE-2021-37679
It is possible to nest a tf.map_fn within another tf.map_fn call. However,
if the input tensor is a RaggedTensor and there is no function signature
provided, code assumes the output is a fully specified tensor and fills output
buffer with uninitialized contents from the heap:
import tensorflow as tf
x = tf.ragged.constant([[1,2,3], [4,5], [6]])
t = tf.map_fn(lambda r: tf.map_fn(lambda y: r, r), x)
z = tf.ragged.constant([[[1,2,3],[1,2,3],[1,2,3]],[[4,5],[4,5]],[[6]]])The t and z outputs should be identical, however this is not the case. The
last row of t contains data from the heap which can be used to leak other
memory information.
The bug lies in the conversion from a Variant tensor to a RaggedTensor. The
implementation
does not check that all inner shapes match and this results in the additional
dimensions in the above example.
The same implementation can result in data loss, if input tensor is tweaked:
import tensorflow as tf
x = tf.ragged.constant([[1,2], [3,4,5], [6]])
t = tf.map_fn(lambda r: tf.map_fn(lambda y: r, r), x)Here, the output tensor will only have 2 elements for each inner dimension.
We have patched the issue in GitHub commit 4e2565483d0ffcadc719bd44893fb7f609bb5f12.
The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.
Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.
This vulnerability has been reported by Haris Sahovic.