DFBScanner: Fast Backdoor Detection in DNNs via Parameter Inspection
Researchers propose DFBScanner, a lightweight framework for detecting backdoor attacks in deep neural networks. Unlike existing methods that rely on activation analysis or trigger reverse engineering—often requiring clean samples or prior knowledge—DFBScanner inspects static parameters in the final classification layer. The key insight is that backdoor-induced feature perturbations cause distinctive parameter updates. This approach dramatically reduces detection time from minutes or hours to milliseconds, matching the speed of advanced attacks. The framework is designed to be practical and generalizable, addressing limitations of current defenses.
Key facts
- DFBScanner is a lightweight static parameter inspection framework.
- It detects backdoor attacks by analyzing parameter updates in the final classification layer.
- Existing defenses are slow, taking minutes or hours, while attacks implant in milliseconds.
- DFBScanner aims to match attack speed with millisecond-level detection.
- The method does not require clean samples or prior knowledge of trigger patterns.
- It shifts focus from recognizing specific trigger patterns to identifying unified backdoor effects.
- The framework is designed for improved efficacy, practicability, and generalizability.
- The research is published on arXiv with ID 2605.18907.
Entities
Institutions
- arXiv