Conceptual

AIM: Additional-Image-Guided Transferable Adversarial Attacks

Adds a plug-and-play Semantic Injection Module to a generator-based attack that pulls target-class semantics from an extra guiding image and injects them into the perturbation, with loss formulations tuned for both targeted and untargeted settings, producing markedly more transferable, target-aligned adversarial examples across surrogate and victim DNNs.