CVE-2021-37677: TensorFlow GPU - Denial of Service via Invalid Arguments in tf.raw_ops.Dequantize

Background

TensorFlow is an end-to-end open source platform for machine learning. It provides a comprehensive ecosystem for building and deploying machine learning models. TensorFlow GPU is a variant of TensorFlow that leverages GPU acceleration for faster computation. It is widely used in various industries and research domains for deep learning tasks.

Vulnerability Detail

The vulnerability (CVE-2021-37677) exists in the shape inference code for tf.raw_ops.Dequantize in TensorFlow GPU. It can be triggered when an attacker provides invalid arguments to the function. The shape inference implementation uses the axis parameter to select between two different values for minmax_rank, which is then used to retrieve tensor dimensions. However, the code assumes that axis can only be -1 or a value greater than -1, without validating other values. This oversight can lead to a segmentation fault (segfault) and result in a denial of service (DoS) condition. The vulnerability affects TensorFlow GPU versions 2.3.0 to 2.3.4, 2.4.0 to 2.4.3, and 2.5.0 to 2.5.1.

pip install --upgrade "tensorflow-gpu==2.12.0"

Workaround

As a temporary workaround, users can avoid the vulnerability by ensuring that valid arguments are provided to the tf.raw_ops.Dequantize function. It is recommended to validate the axis parameter and only pass values of -1 or greater than -1 to avoid triggering the vulnerability. Here's an example of how to validate the axis parameter in Python:

def dequantize_with_validation(tensor, axis):
    if axis < -1:
        raise ValueError('Invalid axis value')
    # Perform dequantization
    result = tf.raw_ops.Dequantize(tensor, axis)
    return result

By implementing this validation check, the vulnerability can be mitigated.

Conclusion

The vulnerability (CVE-2021-37677) in TensorFlow GPU's shape inference code for tf.raw_ops.Dequantize can be exploited by providing invalid arguments, leading to a denial of service (DoS) condition. The TensorFlow development team has addressed the issue in the GitHub commit da857cfa0fde8f79ad0afdbc94e88b5d4bbec764, which will be included in TensorFlow 2.6.0. Users of affected versions (2.3.0 to 2.3.4, 2.4.0 to 2.4.3, and 2.5.0 to 2.5.1) are advised to update their installations to the patched versions once they become available. In the meantime, a temporary workaround is to validate the axis parameter to avoid triggering the vulnerability.

See every vulnerability recorded for tensorflow-gpu →

FAQ