Research map / Label-efficient and robust learning / Recognition and long-tail

Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 193 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FreqNet, a lightweight CNN deepfake detector that learns in the frequency space (high-frequency image/feature representations plus convolutions on FFT amplitude and phase) to generalize to unseen GAN sources trained on ForenSynths ProGAN data.

Problem

Generalizable deepfake image detection: detectors trained on one generator's frequency artifacts overfit and fail on unseen GANs, with limited training data.

Approach

Instead of feeding frequency artifacts to a CNN, the detector itself learns in frequency space: high-pass filtered input, high-frequency filtering of intermediate feature maps across spatial and channel dimensions, and a frequency convolutional layer applied to the amplitude and phase spectra of features.

Key results

• FreqNet with 4-class training reaches 91.5% mean accuracy on the ForenSynths test set, versus 89.1% for Ojha et al.
• Across 17 GAN sources, FreqNet reports 92.8% mean accuracy (+9.8 over Ojha et al.) with 1.9M parameters versus 304M for Ojha et al.
• On the 9 self-synthesized GANs (4-class setting), FreqNet reaches 94.0% mean accuracy versus 77.6% for Ojha et al.

Measured effects reported in the paper

Surprising findings

• FreqNet trained on ProGAN horse/chair/car/cat images transfers to face images, reaching about 98.7% to 99.5% accuracy on ProGAN, StyleGAN and StyleGAN2 faces.

Limitations

• Ablation on the ForenSynths set only; the paper states the ablation will be expanded in revision.
• Performance on some sources is weak, e.g. 69.8% accuracy on the Deepfake set in the 1-class setting.
• Self-synthesized test set and face test set are not publicly described as released.

Experimental setup

Datasets: ForenSynths training set (ProGAN, LSUN) ForenSynths test set self-synthesized GAN images (9 GANs) CelebA-HQ face test set StyleGAN and StyleGAN2 face images

Benchmarks: ForenSynths test set (8 GANs) self-synthesis test set (9 GANs) face test set (ProGAN, StyleGAN, StyleGAN2)

Metrics: accuracy average precision mean accuracy parameter count

Baselines: Wang et al. CNN detector Frank et al. Durall et al. F3Net BiHPF FrePGAN LGrad Ojha et al.

Builds on: F3Net LGrad Ojha et al. universal fake detection ForenSynths

Closest papers by meaning

• FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.91)
• FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.90)
• D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.88)
• FACL-Attack: Frequency-Aware Contrastive Learning for Transferable Adversarial Attacks (AAAI 2024 · similarity 0.87)
• DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.86)
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.86)