# TABOR: A Highly Accurate Approach to Inspecting and Restoring Trojan Backdoors in AI Systems

**Authors:** Wenbo Guo, Lun Wang, Xinyu Xing, Min Du, Dawn Song  
**arXiv:** [1908.01763](https://arxiv.org/abs/1908.01763)  
**Date:** 2019-08-02  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.01763v1/

## Key points

- Trojan backdoor: hidden pattern intentionally implanted in a deep neural network
- Trigger: specific input data that activates the trojan backdoor
- Challenging task to detect and determine trojan backdoors in AI systems
- Existing techniques have limitations and unrealistic assumptions
- TABOR (Trojan Backdoor Inspection and Restoration): new technique for trojan detection
- TABOR approaches trojan detection as a non-convex optimization problem
- TABOR designs a new objective function guided by explainable AI techniques and heuristics
- TABOR defines a novel metric for measuring the quality of identified trojan backdoors
- Anomaly detection method enhances TABOR's ability to identify triggers while filtering out false alarms
- Authors of the research paper: Wenbo Guo, Lun Wang, Xinyu Xing, Min Du, and Dawn Song

## Summary

In the field of artificial intelligence (AI), a trojan backdoor refers to a hidden pattern that is intentionally implanted in a deep neural network. This pattern remains dormant until a specific trigger is present in the input data, causing the infected model to behave abnormally. Detecting and determining the existence of trojan backdoors in AI systems is a challenging task, especially when the triggers vary in size, shape, and position. Previous research has proposed various techniques to address this problem, but none of them fully solve it. These existing techniques often make unrealistic assumptions, such as assuming access to the contaminated training database. Additionally, they struggle to accurately detect trojan backdoors or restore high-fidelity trojan backdoor images. To overcome these limitations, this work introduces TABOR (Trojan Backdoor Inspection and Restoration), a new technique for trojan detection. TABOR approaches trojan detection as a non-convex optimization problem and formulates it as an objective function. Unlike previous methods that also model trojan detection as an optimization problem, TABOR designs a new objective function guided by explainable AI techniques and heuristics. This new objective function enables more effective identification of trojan backdoors. Furthermore, TABOR defines a novel metric for measuring the quality of identified trojan backdoors. By using an anomaly detection method, this metric enhances TABOR's ability to identify intentionally injected triggers in an infected model while filtering out false alarms. The authors of this work are Wenbo Guo, Lun Wang, Xinyu Xing, Min Du, and Dawn Song. Their research paper titled "TABOR: A Highly Accurate Approach to Inspecting and Restoring Trojan Backdoors in AI Systems" presents TABOR as an advanced technique for detecting and addressing trojan backdoors in AI systems with improved accuracy compared to existing methods due its novel objective function and metric for measuring quality of identified backdoors.