CVE-2021-37682: Uninitialized Value Use in TFLite

Background

The TensorFlow Lite (TFLite) package is an open-source deep learning framework developed by TensorFlow. It is designed to run machine learning models on resource-constrained devices such as mobile phones, embedded systems, and microcontrollers. TFLite provides efficient inference for neural networks, including support for quantization to optimize model size and performance. However, a vulnerability has been discovered in TFLite that allows the use of uninitialized values in operations involving quantization.

Vulnerability Detail

The vulnerability in TFLite arises from the improper handling of quantization parameters. In certain parts of the code, the quantization.params field is accessed without proper checks on the corresponding quantization.type field. This can lead to the use of uninitialized values in TFLite operations that involve quantization. For example, the code snippet const auto* affine_quantization = reinterpret_cast<TfLiteAffineQuantization*>(filter->quantization.params); can result in uninitialized values being used if quantization.type is kTfLiteNoQuantization. The impact of this vulnerability is that it can affect the accuracy and reliability of TFLite models that rely on quantization.

pip install --upgrade "tensorflow==2.21.0"

Workaround

As a workaround, users can avoid using quantization in TFLite operations until they have updated to a patched version of TensorFlow. By disabling quantization, the vulnerability can be mitigated. However, this may result in a loss of performance optimization that quantization provides. Here is an example of how to disable quantization in TFLite:

import tensorflow as tf

# Disable quantization
tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT8 = 0

# Rest of the TFLite code
# ...

By setting tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT8 to 0, quantization will be disabled for TFLite operations.

Conclusion

The vulnerability in TFLite that allows the use of uninitialized values in quantization operations has been identified and patched. Users are advised to update their TensorFlow installation to one of the patched versions (2.6.0, 2.5.1, 2.4.3, or 2.3.4) to ensure the security of their TFLite models. A workaround is available by disabling quantization, but this may impact performance optimization. For further support and assistance with vulnerability remediation, please contact DevNack.com.

See every vulnerability recorded for tensorflow →

FAQ