Research on Three-Space Measurement of Backdoor Stealthiness
Start here
An arXiv paper released on October 14, 2026 proposes a three-space measurement framework for backdoor stealth, evaluating representative backdoor attacks in input, feature, and parameter spaces, and testing on four datasets and three model architectures. Results show that stealth is not a single dimension; attacks that are quiet in one space often leave obvious footprints in another, and existing attacks rarely maintain small footprints in all three spaces. The paper also proposes two diagnostic metrics, CDBI and TUP, which characterize class-conditional feature overlap and channels that respond to both triggers and downstream effects, respectively, and states that these metrics are intended to complement, not replace, existing measurement suites.
Generated by AI from reports · Updated 29 days ago
Attention and developments
Attention, hourly
Now 28, 0% in 24 hours·Peak 28 (Nov 6 03:00)
Related research
This event's paper
- Quiet Triggers, Loud Footprints: A Tri-Space Measurement Study of Backdoor Stealthiness
arXiv · 2026-10-14 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningQuiet Triggers, Loud Footprints: A Three-Space Measurement Study of Backdoor Stealthiness
The paper proposes a three-space measurement framework for backdoor stealthiness, evaluating representative backdoor attacks in input, feature, and parameter spaces, and testing on four datasets and three model architectures. Results show that stealthiness is not a single dimension; an attack that is quiet in one space often leaves obvious footprints in another, and existing attacks rarely maintain small footprints in all three spaces. The paper also proposes two diagnostic metrics, CDBI and TUP, which characterize class-conditional feature overlap and channels simultaneously affected by trigger response and downstream impact, respectively, and states that these metrics are intended to complement, not replace, existing measurement suites.