With the rapid advancement of Generative Adversarial Networks (GANs), DeepFake technology has significantly evolved, enabling sophisticated manipulations of facial images and videos. As a result, cyber threats across social media, finance, and politics have increased, prompting the development of proactive disruption techniques. C1-5) However, conventional disruption approaches are inherently sequential, requiring per-image iterative perturbation refinement and resulting in significant computational overhead at scale. To address this challenge, C2-7) we design FlashRush, a parallel execution framework that accelerates proactive DeepFake disruption for large-scale image. FlashRush-I, the first scheme, partitions large-scale image datasets into batches and distributes them across multiple CPU cores and GPUs for parallel processing. FlashRush-P, the second variant, accelerates adversarial perturbation updates by parallelizing computations across hardware resources and merging the updates efficiently. We implement FlashRush within the AntiForgery framework and evaluated it on the Neuron supercomputer using the CelebFaces Attributes (CelebA) dataset against the StarGAN model. Experimental results demonstrate that FlashRush reduces execution time by up to 73.2%, while increasing GPU utilization and memory efficiency by 84.5% and 85.15%, respectively. Moreover, it maintains a competitive adversarial quality, preserving L2 error, SSIM, and PSNR, while achieving a 100% Adversarial Success Rate (ASR). C2-3) These results demonstrate FlashRush as an effective and scalable system for real-time DeepFake disruption.
