WAVIE (Wavelet Augmented Vision Intermediate Embeddings)

Properties

category: detector
maturity: emerging
approach: Lightweight deepfake face detector that projects intermediate transformer embeddings from a frozen CLIP backbone through a learnable module, applies a three-level Daubechies-6 discrete wavelet transform, refines the low-frequency branch while preserving the high-frequency branch, reconstructs the feature via inverse wavelet transform, and classifies the result, aiming for generalization across different GAN and diffusion image-generator families rather than accuracy on a single training distribution.
access: open-source
owns: A small, GPL-3.0-licensed research codebase (GitHub) for cross-generator-generalizable deepfake face detection, trained only on FaceForensics++ and evaluated zero-shot on Celeb-DF-v1, Celeb-DF-v2, and WildDeepFake.
distinctFrom: Unlike intel-fakecatcher (biological PPG blood-flow signal analysis) or the enterprise forensic services (reality-defender, sensity-ai), WAVIE is a small academic codebase whose specific technical contribution is combining a frozen CLIP backbone with wavelet-domain frequency refinement to generalize across unseen GAN/diffusion generators after training on a single dataset, rather than a hosted detection product.
source: arXiv:2609.14437, Lightweight Generalized DeepFake Face Detection with WAVIE
notes: Submitted 2026-09-13 (within window). Code confirmed at github.com/Kappuccino111/WAVIE-Master, GPL-3.0 licensed, 1 star / 1 fork at time of check, single-lab academic code drop; AUROC figures are the authors' own reported results, not independently replicated. Authors: Arya Pulkit, Aditya Ruhela, Akarshan Kapoor, Arnav Bhavsar (IIT Mandi).

Provenance

Source: https://arxiv.org/abs/2609.14437

Properties

category

detector

maturity

emerging

approach

Lightweight deepfake face detector that projects intermediate transformer embeddings from a frozen CLIP backbone through a learnable module, applies a three-level Daubechies-6 discrete wavelet transform, refines the low-frequency branch while preserving the high-frequency branch, reconstructs the feature via inverse wavelet transform, and classifies the result, aiming for generalization across different GAN and diffusion image-generator families rather than accuracy on a single training distribution.

access

open-source

owns

A small, GPL-3.0-licensed research codebase (GitHub) for cross-generator-generalizable deepfake face detection, trained only on FaceForensics++ and evaluated zero-shot on Celeb-DF-v1, Celeb-DF-v2, and WildDeepFake.

distinctFrom

Unlike intel-fakecatcher (biological PPG blood-flow signal analysis) or the enterprise forensic services (reality-defender, sensity-ai), WAVIE is a small academic codebase whose specific technical contribution is combining a frozen CLIP backbone with wavelet-domain frequency refinement to generalize across unseen GAN/diffusion generators after training on a single dataset, rather than a hosted detection product.

source

arXiv:2609.14437, Lightweight Generalized DeepFake Face Detection with WAVIE

notes

Submitted 2026-09-13 (within window). Code confirmed at github.com/Kappuccino111/WAVIE-Master, GPL-3.0 licensed, 1 star / 1 fork at time of check, single-lab academic code drop; AUROC figures are the authors' own reported results, not independently replicated. Authors: Arya Pulkit, Aditya Ruhela, Akarshan Kapoor, Arnav Bhavsar (IIT Mandi).