SCA (Semantic-Consistent Attack) introduces a novel framework for creating photorealistic adversarial examples that effectively fool AI systems while preserving image semantics.

• Achieves 96.5% attack success rate against multiple defenses while maintaining visual quality
• Employs a two-stage attack pipeline: strategic noise injection followed by semantic-preserving refinement
• Demonstrates superior performance compared to existing methods across multiple datasets
• Creates adversarial examples that are challenging to detect even by advanced defensive systems

This research highlights critical security vulnerabilities in neural networks deployed in sensitive environments, presenting concerns for AI systems requiring high reliability and safety guarantees.

Original Paper: SCA: Highly Efficient Semantic-Consistent Unrestricted Adversarial Attack