Executive brief
PaddlePaddle is an open-source deep learning framework used for machine learning model development and deployment. The framework contains an out-of-bounds read vulnerability in the gather_tree function that could allow remote attackers to read sensitive memory data without authentication, potentially exposing training data, model parameters, or other confidential information.
Technical details
An out-of-bounds read vulnerability (CWE-125) exists in the gather_tree function in PaddlePaddle versions prior to 2.4. The vulnerability can be triggered by sending specially crafted input to the affected function without requiring authentication or user interaction, since PaddlePaddle is typically used as a library or service. An attacker can read beyond allocated memory boundaries, potentially disclosing sensitive information from the process memory. The vulnerability was patched in version 2.4, with fixes available in pull request #47051 and subsequent commits that add input validation and error handling.
Affected products
- PaddlePaddle Paddle before 2.4
Timeline
- 2022-12-07: disclosed
- 2022-10-18: patched: Fix available in PR #47051 merged to develop branch