Backdoor attacks against Machine Learning IRL (phrack.org)

🤖 AI Summary
A new vulnerability in the ONNX Runtime (ORT), widely used for deploying deep learning models, has been identified, allowing for backdoor attacks against machine learning models in real-world applications. Researchers discovered a parser flaw that permits ONNX models to exploit arbitrary file and memory reads without validation. This flaw enables the creation of backdoored models that behave normally until prompted with a specific trigger word, at which point they can leak sensitive server data or file contents directly to an attacker. This finding is significant as it simplifies the previously complex techniques required for backdooring models, eliminating the need to tamper with training algorithms and making such attacks more accessible to malicious actors. By embedding logic directly within malicious ONNX files, attackers can initiate data exfiltration through straightforward trigger mechanisms, raising alarms about the security of machine learning deployment practices. The incident underscores the importance of auditing models sourced from public repositories, akin to the necessity of scrutinizing unverified code, as careless implementation can lead to severe breaches of sensitive information.
Loading comments...
loading comments...