CVE-2021-41198: Overflow/crash in tf.tile when tiling tensor is large

Background

The TensorFlow package is an open-source machine learning framework that provides a wide range of tools and libraries for building and deploying machine learning models. It is widely used in various industries and research fields. The tf.tile function in TensorFlow is used to create a new tensor by tiling an existing tensor along specified dimensions. This allows for the replication of the original tensor's values in the new tensor. However, a vulnerability has been discovered in the tf.tile function that can lead to a crash or overflow when the tiling tensor is large.

Vulnerability Detail

The vulnerability arises when the tf.tile function is called with a large input argument. The number of elements in the output tensor becomes too large for the int64_t type, causing an overflow. This overflow is detected by a CHECK statement, resulting in a process crash. An attacker could potentially exploit this vulnerability to crash TensorFlow processes, leading to denial of service. The impact of this vulnerability is considered moderate with a CVSS score of 5.5.

pip install --upgrade "tensorflow==2.21.0"

Workaround

As a workaround, users can avoid calling the tf.tile function with large input arguments. If tiling a large tensor is necessary, it is recommended to split the tiling operation into smaller chunks and perform them separately. This can help prevent the overflow issue and mitigate the vulnerability. Here is an example of how to split the tiling operation:

import tensorflow as tf
import numpy as np

def split_tile(tensor, n, chunk_size):
    tiled_tensor = np.zeros((1, 1, 1))
    for i in range(0, n, chunk_size):
        chunk = tf.keras.backend.tile(x=tensor, n=[min(chunk_size, n - i), 1, 1])
        tiled_tensor = tf.concat([tiled_tensor, chunk], axis=0)
    return tiled_tensor

large_tensor = np.ones((1, 1, 1))
tiled_tensor = split_tile(large_tensor, 100000000, 1000000)

Conclusion

The overflow/crash vulnerability in the tf.tile function of TensorFlow has been addressed with patches in versions 2.6.1, 2.5.2, and 2.4.4. Users are advised to update their installations to these versions to prevent crashes and potential denial of service. If tiling a large tensor is necessary, it is recommended to use the provided workaround of splitting the tiling operation into smaller chunks. For further support and assistance with vulnerability remediation, please contact DevNack.com.

See every vulnerability recorded for tensorflow →

FAQ