Lightweight Generalized DeepFake Face Detection with WAVIE: Wavelet Augmented Vision Intermediate Embeddings
Arya Pulkit, Aditya Ruhela, Akarshan Kapoor, Arnav Bhavsar
cs.CV · cs.AI
Quick verdict
Frozen-CLIP multi-layer fusion plus wavelet refinement achieves strong cross-dataset deepfake AUROC; single-run evidence and mixed baselines temper the claims.
Summary
The paper proposes WAVIE, a lightweight cross-domain deepfake face detector built on a frozen CLIP ViT-L/14. It extracts the [CLS] token from all 24 transformer blocks, projects them through a shared MLP, and fuses them with learned per-layer, per-dimension softmax weights; the fused 1024-d vector undergoes a three-level Daubechies-6 wavelet decomposition in which only the 137-d low-frequency band is refined by a single linear layer before inverse-DWT reconstruction and classification, trained with BCE plus supervised contrastive loss (~2.14M trainable parameters). Trained solely on FaceForensics++ raw frames (~306K), it reports frame-level AUROC of 0.980 in-domain, 0.852 on Celeb-DF-v1, 0.852 on Celeb-DF-v2, and 0.831 on WildDeepfake — the best among the methods it tabulates on Celeb-DF-v2 and WildDeepfake, second to RSG-DA (0.860) on Celeb-DF-v1. Ablations attribute large drops to removing intermediate fusion (0.7532/0.7854) or wavelet refinement (0.7221/0.7636). Main limitations: single-run results with no variance or significance testing; baseline numbers are compiled from original papers under differing protocols, and the closest prior method (RINE) is discussed but not compared; applying a 1-D wavelet across embedding dimensions is conceptually unjustified; evaluation is frame-level only with no compression-robustness results.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fusing intermediate CLIP ViT [CLS] features with learned per-layer weights plus representation-level db6 wavelet low-band refinement achieves 0.852/0.852/0.831 frame AUROC on Celeb-DF-v1/v2/WildDeepfake, best among listed methods on the latter two, with ~2.14M trainable parameters trained only on FF++.
- Both intermediate-layer fusion and wavelet refinement are necessary, since removing either drops CDF AUROC by roughly 7–13 points (Table III).
- Learned per-layer, per-dimension aggregation modestly and consistently outperforms uniform averaging once wavelet refinement is in place.
- Frequency-domain processing at the representation level (not the image level) is a key driver of cross-domain robustness.
Strengths
Precise, internally consistent parameter accounting
Q1 and Q2 (1024×1024 each ≈1.05M), 24×1024 fusion weights (≈24.6K), 137×137 low-band linear (≈18.9K) and the classifier sum to ≈2.14M trainable parameters, matching the claimed budget (Sec. III, Sec. V-A).
Strong cross-dataset numbers under one self-contained training recipe
Trained only on FF++ raw (c0, Table I), WAVIE reaches 0.852/0.852/0.831 frame AUROC on CDF-v1/v2/WDF (Table II) — best among listed methods on CDF-v2 (vs WaViT-CDC 0.816) and WDF (vs 0.812), while keeping the backbone frozen.
Informative ablations quantifying both components
Above-average reproducibility commitment
Honest scoping of limitations
Weaknesses
No statistical rigor anywhere
Mixed-protocol baseline comparison
Missing RINE control
1-D wavelet on embedding dimensions lacks justification
Evaluation scope is narrow
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
WAVIE — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WAVIE in deepfake detection?
WAVIE (Wavelet Augmented Vision Intermediate Embeddings) is a deepfake face detector that runs on a frozen CLIP ViT-L/14, fuses the [CLS] features of all 24 transformer blocks with learned weights, and applies a three-level Daubechies-6 wavelet decomposition to the fused feature, refining only the low-frequency band before classification. Only ~2.14M parameters are trained.
How accurate is WAVIE on unseen deepfake datasets?
Trained only on FaceForensics++, the paper reports frame-level AUROC of 0.980 in-domain, 0.852 on Celeb-DF-v1, 0.852 on Celeb-DF-v2 and 0.831 on WildDeepfake — the best among the methods it tabulates on Celeb-DF-v2 and WildDeepfake, and second to RSG-DA (0.860) on Celeb-DF-v1.
Is the WAVIE code available?
Yes. The authors state they release code, models and preprocessing scripts on GitHub (Kappuccino111/WAVIE-Master), and the paper documents seeds, hyperparameters and preprocessing in enough detail to attempt a retraining.
What are WAVIE's main limitations?
Results come from single runs without variance or significance tests; competitor numbers are compiled from original papers under differing protocols; the closest prior method RINE is discussed but not compared; evaluation is frame-level only, with no compression-robustness or video-level results.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 are present only as captions in the extracted text; the visual claims (t-SNE cluster separation) could not be inspected.
- Table II column alignment is partly inferred from the authors' prose arithmetic; several WDF cells are empty in the extraction.
- The code repository is claimed but was not inspected (offline review).
- No peer-review record is available in the text beyond an IEEE copyright notice; this is arXiv v1.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.