What question did this study set out to answer?

This research aims to develop an effective system for identifying deepfakes, ensuring privacy during inference.

March 25, 2026Open Access

Deep-Check v3: Multi-Dataset Deepfake Detection via EfficientNet-B4 with Multi-Scale Frequency Analysis and Supervised Contrastive Learning

Key Points

This research aims to develop an effective system for identifying deepfakes, ensuring privacy during inference.
Developed Deep-Check v3 architecture combining EfficientNet-B4 and multi-scale frequency analysis.
Trained using focal binary cross-entropy loss and supervised contrastive learning across six public datasets.
Employed test-time augmentation and utilized a lightweight ONNX model for in-browser operations.
Achieved AUC of 1.0000 and EER of 0.04% on the 140K Real vs Fake benchmark.
Demonstrated cross-domain generalization with AUC of 1.0000 on the held-out RVF10K dataset.
The model runs entirely in-browser, enhancing privacy-preserving capabilities.

Abstract

We present Deep-Check v3, a presentation attack detection system achieving AUC=1.0000 and EER=0.04% on the 140K Real vs Fake benchmark (20,000 test images) with Test-Time Augmentation. The architecture combines EfficientNet-B4 (1792-dim) with a novel multi-scale Laplacian frequency branch (FrequencyBranchV2, 48-dim) operating at three spatial scales (32x32, 64x64, 128x128), trained with focal BCE and supervised contrastive learning on six public datasets (~155K images). Cross-domain AUC=1.0000 on the held-out RVF10K dataset demonstrates strong generalization. The 70MB ONNX model runs entirely in-browser via WebAssembly, enabling privacy-preserving inference. Code available at https://github.com/paulrod0/deep-check

Deep-Check v3: Multi-Dataset Deepfake Detection via EfficientNet-B4 with Multi-Scale Frequency Analysis and Supervised Contrastive Learning

Key Points

Abstract

Cite This Study