CVE-2021-41195 : What You Need to Know

TensorFlow experienced a vulnerability in its

tf.math.segment_*

operations, leading to a denial of service due to a

CHECK

failure when processing large segment ids.

Understanding CVE-2021-41195

This CVE highlights a critical vulnerability in TensorFlow versions regarding specific operations that could result in a denial of service attack.

What is CVE-2021-41195?

In TensorFlow versions, a flaw in the implementation of

tf.math.segment_*

operations may lead to a denial of service if a large segment id in

segment_ids

is encountered. This vulnerability is akin to previous reported issues and triggers a

CHECK

failure, causing a service disruption.

The Impact of CVE-2021-41195

The vulnerability's impact is rated as MEDIUM with a base score of 5.5. The attack complexity is low, but the availability impact is high, requiring low privileges for exploitation and having no confidentiality or integrity impact.

Technical Details of CVE-2021-41195

This section delves into the technical aspects of the vulnerability.

Vulnerability Description

The issue arises in the calculation of output shape using

AddDim

, where an overflow of elements in the tensor causes a

CHECK

failure, leading to a

std::abort

. The remedy involves the use of

AddDimWithStatus

instead.

Affected Systems and Versions

TensorFlow version >= 2.6.0 and < 2.6.1
TensorFlow version >= 2.5.0 and < 2.5.2
TensorFlow version < 2.4.4

Exploitation Mechanism

Attack Vector: LOCAL
Privileges Required: LOW
User Interaction: NONE
Scope: UNCHANGED

Mitigation and Prevention

Taking immediate steps and implementing long-term security measures are crucial to mitigate risks effectively.

Immediate Steps to Take

Update TensorFlow to version 2.7.0, or apply the necessary patches if available
Be cautious when processing large segment ids in

segment_ids

Long-Term Security Practices

Regularly update TensorFlow to the latest secure versions
Conduct regular security assessments to identify and address vulnerabilities

Patching and Updates

The fix will be included in TensorFlow 2.7.0, with cherry-picked commits for versions 2.6.1, 2.5.2, and 2.4.4