On Single-Model Transferable Targeted Attacks: A Closer Look at Decision-Level Optimization.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Transferable targeted attacks are crucial in adversarial machine learning but face challenges with existing optimization objectives.
• Common objectives like cross-entropy and Po+Trip loss suffer from gradient vanishing or amplification, limiting attack effectiveness.

Purpose of the Study:

• To investigate intrinsic problems within prevalent optimization objectives for single-model transferable targeted attacks.
• To propose novel methods that mitigate these issues and enhance the transferability of adversarial attacks.

Main Methods:

• Introduced a unified Adversarial Optimization Scheme (AOS) to address gradient vanishing/amplification in cross-entropy and Po+Trip loss.
• Proposed Balanced Logit Loss (BLL) to resolve unbalanced optimization issues in Vanilla Logit Loss (VLL), considering both source and target logits.

Main Results:

• AOS significantly improves targeted transferability by transforming output logits before objective function application.
• BLL effectively addresses unbalanced optimization, preventing source logit increase and boosting transferability.
• Proposed methods demonstrate compatibility and effectiveness across various attack frameworks, datasets (ImageNet, CIFAR-10/100), and challenging scenarios.

Conclusions:

• The developed AOS and BLL offer simple yet potent solutions to enhance adversarial attack transferability.
• These methods provide a robust framework for improving targeted adversarial attacks, even in low-ranked transfer and defense-evasion scenarios.