Related Experiment Video
Updated: Jul 8, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
A two-stage cascaded purification framework for backdoored object detectors via backdoor feature suppression
Lihui Xia1, Lu Zhao2, Junjie Wang1
1College of Computer and Information Engineering, Nanjing Tech University, Nanjing, 211816, China.
Scientific Reports
|July 6, 2026
Summary
This study introduces DAPS-PCMD, a novel two-stage framework to defend object detection models against stealthy backdoor attacks. The method effectively purifies models by severing trigger pathways and decoupling residual features, ensuring robust security and preserving model utility.
Area of Science:
- Computer Vision and Machine Learning
- Cybersecurity for AI Systems
Background:
- Object detectors are crucial for safety-critical applications but vulnerable to backdoor poisoning attacks.
- Existing defenses struggle to balance backdoor suppression and model utility preservation due to complex object detection architectures.
- Backdoor triggers can persist as residual features even after initial pruning, leading to re-activation.
Purpose of the Study:
- To propose a novel two-stage cascaded purification framework, DAPS-PCMD, for robust defense against backdoor attacks in object detection.
- To address the challenge of residual backdoor features and model utility degradation in object detection security.
Main Methods:
- Stage I (DAPS): Utilizes differential activation statistics to identify and prune toxic channels, severing primary trigger pathways and creating a structural prior.
- Stage II (PCMD): Employs prior-constrained residual feature decoupling to suppress persistent trigger associations while maintaining model utility using clean detection anchors.
- The cascaded approach decomposes purification into structural path severing and residual feature decoupling.
Main Results:
- DAPS-PCMD achieved a superior security-utility trade-off across various attack types and strengths.
- Significantly suppressed Attack Success Rate (ASR) to 0.008-0.038 while maintaining high clean mean Average Precision (mAP) at 0.837-0.862.
- Outperformed baseline Robust Neural Network Purification (RNP) with a relative ASR reduction of 78.9%-92.1%.
Conclusions:
- The proposed two-stage cascaded purification framework is key to effectively defending object detection models against sophisticated backdoor attacks.
- DAPS-PCMD offers a robust solution for enhancing the security and reliability of object detection systems in safety-critical domains.
- The explicit decomposition into structural path severing and residual feature decoupling is crucial for achieving optimal defense performance.
Related Concept Videos
01:19Masking and Demasking Agents
EDTA titrations may necessitate masking and demasking agents to temporarily protect a particular metal ion in a mixture from the EDTA reaction. These agents facilitate the sequential analysis of the metal ions by forming stable complexes with some—but not all—metal ions during certain steps.
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on the metal...
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on the metal...
01:05Difference from Background: Limit of Detection
The limit of detection (LOD) is the smallest amount of analyte that can be distinguished from the background noise. The LOD value corresponds to the concentration at which the analyte signal is three times larger than the standard deviation of the blank signal. Below this value, the analyte signal cannot be differentiated from the background noise. It is calculated by dividing the calibration slope by 3 times the standard deviation of the blank signals.
The LOD indicates the presence or absence...
The LOD indicates the presence or absence...