Adversarial Attack and Defense in Deep Ranking.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are susceptible to adversarial attacks, causing misclassifications.
• The vulnerability of DNN-based image ranking systems to such attacks is not well-understood.

Purpose of the Study:

• To propose novel adversarial attacks (Candidate Attack, Query Attack) targeting deep ranking systems.
• To develop an effective defense mechanism (anti-collapse triplet defense) to enhance model robustness.

Main Methods:

• Representing ranking order as inequalities and designing a triplet-like objective function for optimal perturbation.
• Developing an anti-collapse triplet defense to prevent positive and negative sample collapse.
• Proposing an empirical robustness score to evaluate defense effectiveness against diverse attacks.

Main Results:

• Demonstrated that proposed attacks can effectively compromise deep ranking systems.
• Showcased that the proposed defense significantly improves ranking system robustness.
• Validated effectiveness across multiple datasets (MNIST, Fashion-MNIST, CUB200-2011, CARS196, Stanford Online Products).

Conclusions:

• Deep ranking systems are vulnerable to specific adversarial attacks.
• The proposed defense strategy enhances robustness and mitigates a broad spectrum of attacks.
• This work contributes to securing DNN-based image ranking applications.