Adversarial Examples Generation for Deep Product Quantization Networks on Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Product Quantization Networks (DPQNs) excel at image retrieval due to efficient high-dimensional feature encoding.
• Deep Neural Networks (DNNs), including DPQNs, are susceptible to adversarial examples—subtle perturbations designed to cause misclassification or retrieval errors.
• The vulnerability of DPQNs to adversarial attacks in retrieval systems poses a significant safety concern for commercial applications.

Purpose of the Study:

• To investigate the impact of adversarial examples on DPQN-based image retrieval systems.
• To propose a novel framework for generating adversarial query images to attack DPQNs.
• To extend existing adversarial attack methods to DPQN retrieval scenarios.

Main Methods:

• Developed an adversarial example generation framework targeting DPQN retrieval systems.
• Perturbed the probability distribution of centroid assignments for clean queries, bypassing reliance on ground-truth labels.
• Implemented both white-box and black-box non-targeted attacks.
• Extended non-targeted attacks to targeted attacks using a novel sample space averaging scheme (AS).

Main Results:

• Successfully generated adversarial examples that effectively mislead target DPQNs.
• Demonstrated significant degradation of retrieval performance across diverse experimental settings.
• Validated the effectiveness of both non-targeted and targeted adversarial attacks.
• Achieved theoretical guarantees for the proposed targeted attack method.

Conclusions:

• Adversarial attacks pose a tangible threat to the reliability of DPQN-based image retrieval systems.
• The proposed framework provides a viable method for generating adversarial examples against DPQNs.
• Further research is needed to develop robust defense mechanisms for DPQN retrieval systems against such attacks.