DIWA-Net

What it is

DIWA-Net is a dual-stream architecture for open-set multimodal deepfake detection. Video and audio run on frozen foundation backbones — DINOv2 (ViT-S/14) and Wav2Vec 2.0 — adapted with LoRA. Final-year BS CS thesis at UET Taxila, supervised by Dr. Rabbia Mahum (UET Taxila; Researcher at KFUPM).

Architecture

• Two frozen foundation backbones adapted with LoRA (rank r=16, alpha=32) in the attention projections.
• About 4.2M trainable parameters out of 120M total (~3.5%).
• Task heads warm-started before adapters were added.

Why LoRA

Full fine-tuning left no usable learning rate. Above 1e-5, validation AUC collapsed while training AUC kept climbing (representation drift). Below 1e-5, training was stable but almost useless. Constraining backbone updates to low-rank subspaces allowed a much higher learning rate without that instability.

Results

Evaluated on MAVOS-DD with an identity-isolated split:

• 0.9952 in-domain AUC
• 0.9592 open-set AUC (unseen generation methods and languages at test time)

The open-set number is the one that matters. Before trusting these results, two published baselines were reproduced under the same protocol and matched within 0.4 AUC points.

Publication

"DIWA-Net: Dual-Stream Multimodal Deepfake Detection with Parameter-Efficient Adaptation." Accepted and presented at IBCAST 2026. Journal extension in progress.