CVE-2021-29589: TensorFlow-GPU - Division by Zero Error in GatherNd TFLite Operator

Background

TensorFlow is an end-to-end open source platform for machine learning. It provides a comprehensive ecosystem of tools, libraries, and resources for developing and deploying machine learning models. TensorFlow-GPU is a variant of TensorFlow that leverages GPU acceleration for faster computation. The GatherNd TFLite operator is a key component in TensorFlow-GPU that allows gathering elements from multiple tensors based on indices. This operator is vulnerable to a division by zero error when the 'params' input tensor is empty, leading to potential crashes or incorrect results in machine learning models.

Vulnerability Detail

The vulnerability in the GatherNd TFLite operator arises from a division by zero error in the reference implementation code. When the 'params' input tensor is empty, the 'params_shape.Dims(.)' function returns zero for at least one dimension. This triggers a division by zero error, which can have various consequences depending on how the operator is used within a machine learning model. An attacker can exploit this vulnerability by crafting a malicious model that triggers the division by zero error, potentially leading to crashes, incorrect results, or even remote code execution. The vulnerability has been assigned the identifier CVE-2021-29589.

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

Workaround

As a temporary workaround, you can modify the affected machine learning models to handle the case of an empty 'params' input tensor. One possible approach is to add a check for an empty tensor before invoking the GatherNd TFLite operator. If the 'params' tensor is empty, you can either skip the operation or provide default values to avoid triggering the division by zero error. Here's an example of how you can implement this workaround in Python:

import tensorflow as tf

params = ...  # Your 'params' tensor

if tf.size(params) == 0:
    # Handle the empty tensor case
    ...
else:
    # Invoke the GatherNd TFLite operator
    output = tf.gather_nd(params, indices)

Note that this workaround is not a permanent solution and should be replaced with the recommended fix as soon as possible.

Conclusion

The vulnerability in the GatherNd TFLite operator of TensorFlow-GPU can result in a division by zero error when the 'params' input tensor is empty. This can lead to crashes, incorrect results, or even remote code execution. It is crucial to upgrade to the latest version of TensorFlow-GPU or apply the corresponding fixes to mitigate the risk. Rebuilding and retesting affected machine learning models is also necessary to ensure correct behavior. For further support on vulnerability remediation, please contact DevNack.com.

See every vulnerability recorded for tensorflow-gpu →

FAQ