tfsa-2022-007.md

TFSA-2022-007: Type confusion in shape inference for ConcatV2

CVE Number

CVE-2022-21731

Impact

The implementation of shape inference for ConcatV2 can be used to trigger a denial of service attack via a segfault caused by a type confusion:

import tensorflow as tf

@tf.function
def test():
  y = tf.raw_ops.ConcatV2(
    values=[[1,2,3],[4,5,6]],
    axis = 0xb500005b)
  return y

test()

The axis argument is translated into concat_dim in the ConcatShapeHelper helper function. Then, a value for min_rank is computed based on concat_dim. This is then used to validate that the values tensor has at least the required rank:

  int64_t concat_dim;
  if (concat_dim_t->dtype() == DT_INT32) {
    concat_dim = static_cast<int64_t>(concat_dim_t->flat<int32>()(0));
  } else {
    concat_dim = concat_dim_t->flat<int64_t>()(0);
  }

  // Minimum required number of dimensions.
  const int min_rank = concat_dim < 0 ? -concat_dim : concat_dim + 1;

  // ...
  ShapeHandle input = c->input(end_value_index - 1);
  TF_RETURN_IF_ERROR(c->WithRankAtLeast(input, min_rank, &input));

However, WithRankAtLeast receives the lower bound as a 64-bits value and then compares it against the maximum 32-bits integer value that could be represented:

Status InferenceContext::WithRankAtLeast(ShapeHandle shape, int64_t rank,
                                         ShapeHandle* out) {
  if (rank > kint32max) {
    return errors::InvalidArgument("Rank cannot exceed kint32max");
  }
  // ...
}

Due to the fact that min_rank is a 32-bits value and the value of axis, the rank argument is a negative value, so the error check is bypassed.

Patches

We have patched the issue in GitHub commit 08d7b00c0a5a20926363849f611729f53f3ec022.

The fix will be included in TensorFlow 2.8.0. We will also cherrypick this commit on TensorFlow 2.7.1, TensorFlow 2.6.3, and TensorFlow 2.5.3, 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 Yu Tian of Qihoo 360 AIVul Team.