Stealthy and efficient adversarial example attack on video retrieval systems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video-focused social media drives demand for efficient video retrieval systems.
• Machine learning powers these systems but is vulnerable to adversarial example (AE) attacks.
• Existing AE attacks on videos are often detectable due to widespread perturbations.

Purpose of the Study:

• To propose DUO, a stealthy, targeted, black-box AE attack for video retrieval systems.
• To enhance stealthiness and query efficiency in AE attacks.
• To develop sparse perturbation generation methods for improved AE attack performance.

Main Methods:

• DUO utilizes a DUal search Over frame-pixel strategy for sparse perturbations.
• Key observations include the importance of key video frames, differential pixel/frame contribution, and pixel locality.
• Two variants, DUO_P (pixel sparsity) and DUO_G (group sparsity), are proposed using SparseTransfer and SparseQuery components.

Main Results:

• DUO demonstrates improved stealthiness and efficacy compared to existing AE attacks.
• DUO_P achieves higher precision and reduces adversarial perturbations by over 100x.
• DUO_G requires over 10x fewer queries than state-of-the-art methods.

Conclusions:

• DUO offers a significant advancement in stealthy and efficient adversarial attacks against video retrieval systems.
• The proposed sparsity-based methods effectively reduce perturbations and query counts.
• DUO provides a robust framework for evaluating and improving the security of video retrieval technologies.