Generating Semantic Adversarial Examples via Feature Manipulation in Latent Space.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks, where subtle perturbations mislead classifiers.
• Traditional attacks use unstructured pixel changes, often altering image appearance unnaturally and lacking interpretability.

Purpose of the Study:

• To develop a new adversarial attack strategy focusing on semantic, structural image perturbations.
• To enhance the interpretability and effectiveness of adversarial examples against DNNs.

Main Methods:

• Proposed a methodology manipulating semantic attributes via disentangled latent codes.
• Engineered adversarial perturbations by altering single or combined latent codes.
• Introduced two unsupervised semantic manipulation strategies: vector-disentangled and feature map-disentangled representations.

Main Results:

• Demonstrated the attack's potency on real-world image data, especially against black-box classifiers.
• Empirical evaluations confirmed the effectiveness of semantic-oriented structural perturbations.
• Established the existence of a universal semantic adversarial example, independent of specific training data.

Conclusions:

• The proposed semantic manipulation attack is a potent threat to DNNs, particularly in black-box scenarios.
• Semantic adversarial examples offer a more interpretable and potentially more powerful attack vector.
• The discovery of universal semantic adversarial examples has significant implications for AI security research.