Related Experiment Video
Updated: Aug 14, 2026
09:49
Holistic Facial Composite Creation and Subsequent Video Line-up Eyewitness Identification Paradigm
Published on: December 24, 2015
Adversarial face camouflage based on multi-parameter enhancement
DaPeng Men1, JingYu Wang1, XiaoLin Zhang1
1School of Digital and Intelligent Industry, Inner Mongolia University of Science and Technology, No. 7 Aerding Street, Kundulun District, Baotou, 014010, Inner Mongolia, China.
Summary
Researchers developed a new facial recognition attack method to improve adversarial attack transferability. This Multi-Initialization Enhanced Aggregation (MEA) method enhances surrogate model diversity, boosting attack effectiveness against facial recognition systems.
Area of Science:
- Computer Science
- Artificial Intelligence
- Cybersecurity
Background:
- Facial recognition (FR) models are susceptible to adversarial attacks, where manipulated images exploit system vulnerabilities.
- Current methods for improving adversarial attack transferability are limited by insufficient use of diverse initialization strategies for surrogate models.
Purpose of the Study:
- To propose a novel attack method, Multi-Initialization Enhanced Aggregation (MEA), to enhance the transferability of adversarial attacks in facial recognition.
- To address the limitations of existing methods in leveraging diverse initialization strategies for surrogate models.
Main Methods:
- The MEA attack method comprises two stages: Multi-Initialization Adversarial Enhancement (MIAE) and Enhanced Adversarial Aggregation (EAA).
- MIAE enhances surrogate model diversity using checkpoint saving based on diversity metrics and multi-layer initialization.
- EAA improves transferability by perturbing high-level features and incorporating an adversarial makeup technique for disguise generation.
Main Results:
- The MEA method demonstrated superior performance compared to existing facial adversarial attacks.
- MEA achieved a 20.35% improvement over the second-best input transformation attack.
- The proposed method showed a 9.22% improvement over current facial adversarial attack techniques.
Conclusions:
- The Multi-Initialization Enhanced Aggregation (MEA) attack method significantly enhances the transferability of adversarial attacks against facial recognition systems.
- MEA's novel approach, combining diverse initializations and feature perturbations, offers a more effective strategy for evaluating FR model security.