FFCBA: Feature-based Full-target Clean-label Backdoor Attacks
Yangxu Yin, Honglong Chen, Yudong Gao, Peng Sun, Liantao Wu, Zhe Li, Weifeng Liu·April 29, 2025
Summary
FFCBA introduces feature-based clean-label multi-target backdoor attacks, proposing Feature-Spanning and Feature-Migrating Backdoor Attacks. These methods generate noise triggers aligned with original features, enhancing effectiveness and specificity. FMBA specifically boosts cross-model attack capability. FFCBA surpasses state-of-the-art defenses, demonstrating robust performance across multiple datasets and models.
Introduction
Background
Overview of backdoor attacks in machine learning
Importance of clean-label backdoor attacks
Objective
To introduce and evaluate FFCBA, a novel approach to feature-based clean-label multi-target backdoor attacks
Method
Feature-Spanning Backdoor Attack
Explanation of the attack mechanism
Data collection methods for crafting feature-spanning triggers
Data preprocessing techniques for enhancing trigger effectiveness
Feature-Migrating Backdoor Attack
Description of the attack's unique approach
Data collection and preprocessing for feature-migrating triggers
FMBA: Feature-Migrating Backdoor Attack
Focus on FMBA's capability to boost cross-model attack effectiveness
Detailed explanation of FMBA's implementation and benefits
Results
Performance Evaluation
Comparison with state-of-the-art backdoor attack methods
Metrics used for assessing attack effectiveness
Robustness Across Datasets and Models
Demonstration of FFCBA's performance across various datasets
Analysis of FFCBA's adaptability to different machine learning models
Conclusion
Summary of Findings
Recap of FFCBA's contributions to the field of backdoor attacks
Future Work
Suggestions for further research and improvements
Implications
Discussion on the ethical and practical implications of FFCBA
Basic info
papers
cryptography and security
artificial intelligence
Advanced features
Insights
How do Feature-Spanning and Feature-Migrating Backdoor Attacks differ in their approach to generating noise triggers?
In what ways does FFCBA demonstrate robust performance across different datasets and models?
What are the key steps involved in implementing FMBA to enhance cross-model attack capability?
What innovative techniques does FFCBA introduce to surpass state-of-the-art defenses?