A Unified Framework for Backdoor Trigger Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep Neural Networks (DNNs) are vulnerable to backdoor attacks, where hidden triggers manipulate model behavior.
• Stealthy triggers are difficult to detect visually, posing significant security challenges.
• Identifying these triggers is crucial for attack analysis and defense strategy development.

Purpose of the Study:

• To propose the Backdoor Trigger Segmentation (BTS) task for identifying hidden triggers in DNNs.
• To establish a comprehensive benchmark for evaluating BTS methods.
• To develop a unified algorithm for segmenting diverse backdoor triggers.

Main Methods:

• Developed a mathematical framework to unify various backdoor triggers.
• Introduced a unified Trigger Locator (TriLoc) algorithm for backdoor trigger segmentation.
• Created a benchmark dataset with 8 attack methods, 8 triggers, and 179 settings for image and text data.

Main Results:

• The proposed TriLoc algorithm demonstrated superior performance in segmenting backdoor triggers compared to state-of-the-art methods.
• Extensive experiments validated the effectiveness of TriLoc across various attack settings and data modalities.
• The benchmark provides a standardized evaluation platform for BTS research.

Conclusions:

• Backdoor Trigger Segmentation (BTS) is a critical task for addressing security threats in DNNs.
• The proposed TriLoc algorithm offers a robust and unified approach to segmenting diverse backdoor triggers.
• The BTS benchmark facilitates further research and development of defense mechanisms against backdoor attacks.