arxiv_cv 95% Match Research Paper AI security researchers,Computer vision engineers,Developers of safety-critical AI systems 3 weeks ago

Structured Universal Adversarial Attacks on Object Detection for Video Sequences

computer-vision › object-detection

📄 Abstract

Abstract: Video-based object detection plays a vital role in safety-critical applications. While deep learning-based object detectors have achieved impressive performance, they remain vulnerable to adversarial attacks, particularly those involving universal perturbations. In this work, we propose a minimally distorted universal adversarial attack tailored for video object detection, which leverages nuclear norm regularization to promote structured perturbations concentrated in the background. To optimize this formulation efficiently, we employ an adaptive, optimistic exponentiated gradient method that enhances both scalability and convergence. Our results demonstrate that the proposed attack outperforms both low-rank projected gradient descent and Frank-Wolfe based attacks in effectiveness while maintaining high stealthiness. All code and data are publicly available at https://github.com/jsve96/AO-Exp-Attack.

Authors (3)

Sven Jacob

Weijia Shao

Gjergji Kasneci

Submitted

October 16, 2025

arXiv Category

cs.CV

arXiv PDF Code

Key Contributions

This paper proposes a novel, minimally distorted universal adversarial attack specifically for video object detection. It leverages nuclear norm regularization to create structured, background-focused perturbations and employs an adaptive gradient method for efficient optimization, outperforming existing attacks in effectiveness and stealthiness.

Business Value

Enhances the understanding of security vulnerabilities in critical AI systems like autonomous vehicles and surveillance, driving the development of more robust detection models.

Paper Metadata

Innovation Type

Algorithmic Improvement

Deployment Feasibility

High for research and security testing; the attack itself is a tool for evaluation.

Limitations Addressed

Vulnerability of deep learning object detectors to adversarial attacks,Ineffectiveness of existing universal attacks in video sequences,Lack of stealthiness in adversarial perturbations

Performance Gains

Outperforms low-rank projected gradient descent and Frank-Wolfe based attacks in effectiveness while maintaining high stealthiness.

View Code on GitHub

Technical Tags

adversarial attacksobject detectionvideo sequencesuniversal perturbationsnuclear norm regularizationstructured perturbationsbackground attacksadaptive gradient methodstealthinessminimally distorted

Research Topics

Adversarial RobustnessComputer Vision SecurityObject DetectionVideo AnalysisMachine Learning Security

Methods & Architectures

Nuclear Norm RegularizationAdaptive Optimistic Exponentiated Gradient MethodMinimally Distorted Universal Adversarial Attack Deep Learning Object Detectors

Applications & Tasks

Autonomous Driving Surveillance Robotics Safety-Critical Systems Adversarial VulnerabilityAttack EffectivenessAttack Stealthiness Adversarial Attack GenerationObject Detection Evaluation

Related Fields

Computer VisionMachine Learning SecurityDeep LearningCybersecurity

Keywords

Adversarial AttacksObject DetectionVideoUniversal PerturbationsNuclear NormStealthy AttacksDeep Learning SecurityRobustnessComputer VisionCybersecurityGradient Methods

Academic Context

#Adversarial Robustness#Computer Vision Security#Object Detection#Video Analysis#Machine Learning Security

Commercial Potential

Potential Products

Adversarial attack simulation toolsRobust object detection model benchmarks

Target Industries

AutomotiveAerospaceSecurity and SurveillanceRobotics

Use Case Examples

Testing the robustness of self-driving car perception systemsEvaluating surveillance camera AI against manipulationDeveloping defenses against adversarial manipulation

Competitive Edge

The proposed attack method is positioned as superior to existing low-rank projected gradient descent and Frank-Wolfe based attacks in terms of effectiveness and stealthiness for video object detection.

Market Opportunity

Growing market for AI security and robustness testing.

Revenue Models

Consulting services for AI security audits.

Resource Requirements

Compute Needs

Moderate to high, depending on the complexity of the target object detector and video sequence length.

Data Requirements

Video datasets with object annotations.

Deployment Constraints

The attack is for evaluation, not direct deployment in a production system.

Scalability

The adaptive gradient method aims to improve scalability.

Production Readiness

Maturity Level

Research

Time to Market

N/A (research tool)

Licensing

Publicly available code implies permissive licensing.

Patent Potential

Low, as it describes an attack method.

View Full Paper Back to Papers