We present DÆLTA, a Dual-branch AutoEncoder Latent Transformation Attack framework designed to generate adversarial versions of deepfake images in a transfer black-box setting, avoiding the need to optimize perturbations against a specific detector. DÆLTA projects each input through an autoencoder trained exclusively on a curated subset of real Flickr-Faces-HQ images. The architecture behind DÆLTA encodes each image via two parallel branches: one operating on RGB pixels and the other on log-scaled RGB Discrete Cosine Transform (DCT) representations augmented with self-attention. The resulting 256-dimensional latent feature vectors are concatenated and decoded by a shared transposed-convolutional decoder. Training is designed to minimize a composite perceptual loss combining different quality metrics following the guidelines of the ACM Multimedia Fake Face Detection Challenge (AADD 2026), on whose benchmark we evaluate DÆLTA against 2 white-box and 8 black-box detectors with heterogeneous architectures. Our method demonstrates cross-model transferability, with an Attack Success Rate (ASR) above 86% on six out of eight black-box detectors, outperforming many gradient-based baselines.
Manifold-Guided Adversarial Attack of Deepfake Detectors via Dual-Branch Autoencoder Projection
Cirillo, Stefano;Di Biasi, Luigi;Genito, Giuseppe;Solimando, Giandomenico
2026
Abstract
We present DÆLTA, a Dual-branch AutoEncoder Latent Transformation Attack framework designed to generate adversarial versions of deepfake images in a transfer black-box setting, avoiding the need to optimize perturbations against a specific detector. DÆLTA projects each input through an autoencoder trained exclusively on a curated subset of real Flickr-Faces-HQ images. The architecture behind DÆLTA encodes each image via two parallel branches: one operating on RGB pixels and the other on log-scaled RGB Discrete Cosine Transform (DCT) representations augmented with self-attention. The resulting 256-dimensional latent feature vectors are concatenated and decoded by a shared transposed-convolutional decoder. Training is designed to minimize a composite perceptual loss combining different quality metrics following the guidelines of the ACM Multimedia Fake Face Detection Challenge (AADD 2026), on whose benchmark we evaluate DÆLTA against 2 white-box and 8 black-box detectors with heterogeneous architectures. Our method demonstrates cross-model transferability, with an Attack Success Rate (ASR) above 86% on six out of eight black-box detectors, outperforming many gradient-based baselines.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


