CVE-2023-25670: TensorFlow CPU Null Pointer Error in QuantizedMatMulWithBiasAndDequantize

Background

TensorFlow is an open-source machine learning framework widely used for building and training deep learning models. It provides a comprehensive ecosystem for developing applications that require complex mathematical computations. The TensorFlow CPU package is specifically designed for environments where GPU acceleration is not available. This package includes various operations, including quantized matrix multiplication, which optimizes performance for low-precision data types. However, a recent vulnerability has been identified in the QuantizedMatMulWithBiasAndDequantize operation, which can lead to a null pointer error under certain conditions.

Vulnerability Detail

The vulnerability arises from a null pointer error (NPE) in the QuantizedMatMulWithBiasAndDequantize function when using the MKL (Math Kernel Library) feature. This can occur when specific parameters are passed to the function, leading to potential crashes or unexpected behavior in applications utilizing this operation. The issue is particularly concerning as it may affect the stability and reliability of machine learning models that rely on this functionality, potentially causing data loss or corruption during processing.

pip install --upgrade "tensorflow-cpu==2.21.0"

pip install --upgrade "tensorflow-cpu==2.11.1"

Workaround

As a temporary workaround, you can avoid using the QuantizedMatMulWithBiasAndDequantize function with the specific parameters that trigger the null pointer error. Here is an example of how to modify the parameters to prevent the error:

import tensorflow as tf

func = tf.raw_ops.QuantizedMatMulWithBiasAndDequantize
para = {
    'a': tf.constant(138, dtype=tf.quint8),
    'b': tf.constant(4, dtype=tf.qint8),
    'bias': [[31.81644630432129, 47.21876525878906], [109.95201110839844, 152.07968139648438]],
    'min_a': 141.5337138686371,
    'max_a': [73.84139251708984, 173.15280151367188],
    'min_b': [],
    'max_b': [[16.128345489501953, 193.26820373535156]],
    'min_freezed_output': [],
    'max_freezed_output': [115.50032806396484, 156.974853515625],
    'Toutput': 1.0,
    'transpose_a': True,
    'transpose_b': False,
    'input_quant_mode': 'MIN_FIRST'
}

# Modify parameters to avoid NPE
# Ensure min_b and min_freezed_output are set appropriately
func(**para)

Conclusion

The identified vulnerability in TensorFlow CPU's QuantizedMatMulWithBiasAndDequantize function poses a risk of null pointer errors, which can affect application stability. It is crucial to update to the patched version 2.11.1 or later to mitigate this risk. Additionally, users can implement workarounds to avoid triggering the error until they can update.

See every vulnerability recorded for tensorflow-cpu →

FAQ