cs.CVSep 30, 2026
SaveA Generalizable and Explainable Framework for Synthetic Video Detection Using First-Digit Gradient Statistics
Organizations: B.Tech CSE IIT Jodhpur, India · School of Automation & Robotics GGSIPU, New Delhi, India · Centre for Artificial Intelligence MITS Gwalior, India
Abstract
AI video generators have not only become harder to detect but are used to generate a diverse set of scenarios from landscapes to street views to animal videos. This creates a problem where CNN-based detectors are effective but offer no insight into their inner workings, while forensics-based detectors are often pretrained for a set scenario or become too complex to derive meaningful insights. We present a novel approach to AI video detection using Sobel gradient values analysed with the first-digit law. Using linear discriminant analysis, we visualise the discriminatory signal, while a multi-layer perceptron is used for classification. The detection method has no generator- or scenespecific features, and the model has no knowledge of container formats, codec, bitrate, or compression artefacts. The model is trained and tested on GenBuster-200K, GenBusterBench, GenVA, FaceForensics++ C23, and CelebDF. We also show how zero-shot detection fails even though the feature set carries a discriminatory signal.
Figures & tables
Fig. 1: A Generalizable and Explainable Framework for Synthetic Video Detection Using First-Digit Gradient Statistics. Each video is resampled to a uniform frame rate, split into spatial ( ) and temporal-derivative ( ) banks, and passed through Sobel gradient computation over ten per-frame channels. The resulting gradient responses are summarised as first-digit histograms and energy terms, concatenated into an 800-D feature vector, and either projected with LDA for visualisation or classified directly by the MLP to yield a real/synthetic decision.
| Symbol | Definition | Rationale |
|---|---|---|
| Raw BGR colour planes | Per-channel edge structure | |
| Luminance baseline | ||
| Sharpness / focus energy | ||
| Channel-correlation check | ||
| Channel-correlation check | ||
| Channel-correlation check |
TABLE I: The ten feature channels derived from each frame.
Fig. 2: Representative frames from GenBuster-200K. From left to right: real, CogVideoX, EasyAnimate, HunyuanVideo, LTXVideo.
Fig. 3: Representative frame pair from CelebDF. Left: real (Celeb-real). Right: Celeb-synthesis face-swap fake of the same subject. The manipulation is visually near-imperceptible, consistent with the low 0.624 AUC reported in Section V-E .
Fig. 4: LDA on the GenBuster-200K test split. Left: LD1 kernel-density estimate per generator. Right: LD1 vs. LD2 scatter with class centroids ( ).
| Generator | Det. rate | AUC | |
|---|---|---|---|
| CogVideoX | 534 | 93.1% | 0.988 |
| EasyAnimate | 733 | 99.6% | 0.998 |
| Gen3 | 32 | 50.0% | 0.922 |
| HunyuanVideo | 202 | 100.0% | 1.000 |
| JiMeng | 31 | 80.6% | 0.965 |
| Kling | 34 | 64.7% | 0.954 |
TABLE II: Per-generator detection on the GenBuster-200K in-domain test set.
| Metric | Value |
|---|---|
| ROC-AUC | 0.986 |
| Balanced Acc. | 0.945 |
| Real recall | 97.1% |
| Fake recall | 91.9% |
TABLE III: GenBuster-200K in-domain binary detection.
Fig. 5: GenBuster-200K MLP training curves (loss, balanced accuracy, and ROC-AUC per epoch on the validation split).
Fig. 6: Cohen’s on GenBuster-200K. Left: top-20 features by (blue = synthetic real, red = synthetic real). Right: cumulative discriminative power; crosshairs at 50%, 80%, 95%.
Fig. 7: LDA on GenBusterBench (all generators). Multiple per-generator clusters are visible in the scatter, each deviating from the real centroid in a distinct direction—the geometric signature of generator-specific artefact distributions.
Fig. 8: LDA on GenBuster-200K features coloured by distribution group (real, in-distribution fakes, benchmark fakes, test fakes). In-distribution and benchmark fakes deviate from the real class in different LD directions.
| Tier | AUC | BalAcc | Real Rec. | Fake Rec. |
|---|---|---|---|---|
| Standard | 0.976 | 0.938 | 93.0% | 94.5% |
| Extended | 0.984 | 0.950 | 93.0% | 97.0% |
| Wild (newest) | 0.939 | 0.898 | 93.0% | 86.7% |
TABLE IV: GenBusterBench results by evaluation tier.
Fig. 9: Cohen’s on GenBusterBench. All top-20 features are negative (red); effect sizes are uniformly large ( ), with energy features dominating.
Fig. 10: Training curves for the GenBusterBench MLP (loss, balanced accuracy, and ROC-AUC per epoch).
| Generator | Det. rate | |
|---|---|---|
| Pika v1 | 5,447 | 99.7% |
| Sora | 5,452 | 99.2% |
| VideoCrafter-2 | 5,452 | 97.1% |
| Overall | 16,351 | AUC 0.993 / BalAcc 0.979 |
TABLE V: GenVA zero-shot detection results (no fine-tuning).
Fig. 11: LDA on GenVA fakes combined with real videos from the GenBuster-200K test split. All three GenVA generators form tight clusters clearly separated from the real class, despite the model never having been trained on GenVA.
Fig. 12: Cohen’s on GenVA. Effect sizes are extreme ( up to ); the steep cumulative curve indicates the signal is concentrated in a small feature subset, consistent with near-perfect zero-shot detection.
Fig. 13: LDA on FaceForensics++ restricted to fake classes only. The five forgery methods form distinct clusters, yet their joint separation from the real class (cf. Fig. 14 ) is limited—consistent with the pixel-mass hypothesis: stitching-based forgeries transplant real pixels rather than synthesizing new gradient distributions.
Fig. 14: LDA vs. PLS-DA on FaceForensics++ (six classes). Rows from top: LDA distribution / scatter, PLS-DA distribution / scatter, PLS2 distribution and LD1-vs-PLS2 cross-scatter. The marginal separation under both projections (LDA BalAcc = 0.521) confirms limited discriminability of face-swap artefacts in the first-digit gradient feature space.
| Metric | Value |
|---|---|
| ROC-AUC (binary) | 0.717 |
| Balanced Acc. | 0.596 |
| Real recall | 29.5% |
| Fake recall | 89.6% |
TABLE VI: FaceForensics++ collapsed binary metrics (real vs. all fakes).
| Class | Recall |
|---|---|
| Deepfakes | 43.0% |
| Face2Face | 50.5% |
| FaceShifter | 77.5% |
| FaceSwap | 56.0% |
| NeuralTextures | 58.0% |
| DeepFakeDetection ∗ | 79.4% |
TABLE VII: FaceForensics++ per-class recall.
Fig. 16: Cohen’s on FaceForensics++ (original vs. all manipulated pooled, 1,600-D face+background). Effect sizes are uniformly small ( ); the nearly linear cumulative curve is the quantitative signature of the pixel-mass limitation.
Fig. 17: CelebDF MLP training curves (loss, test accuracy/balanced accuracy, and test ROC-AUC per epoch; full 800-D feature, 80/20 split).
| Metric | Value |
|---|---|
| ROC-AUC | 0.624 |
| Balanced Acc. | 0.615 |
| Real recall | 24.2% |
| Fake recall | 98.8% |
TABLE VIII: CelebDF binary detection (full 800-D feature, 80/20 split).
| Comparison | Pairs | mean | mean JSD |
| Cross-dataset, real class | 6 | 0.959 | 0.093 |
| Cross-dataset, pooled fake class | 10 | 0.961 | 0.081 |
| Cross-generator, GB200K (12 gens.) | 66 | 0.993 | 0.038 |
| Cross-generator, GenBusterBench (12 gens.) | 66 | 0.994 | 0.038 |
| Cross-generator, GenVA (3 gens.) | 3 | 0.987 | 0.059 |
| Cross-generator, FaceForensics++ (6 cls.) | 15 | 0.990 | 0.025 |
TABLE IX: Empirical-vs-empirical FSD histogram stability (not compared to the theoretical Benford curve). Averaged over all 80 modes.
Fig. 18: Representative first-significant-digit distributions vs. the theoretical Benford curve (temporal-derivative, , both gradient directions) for all five datasets: (a) GenBuster-200K, (b) GenBusterBench, (c) GenVA (fakes combined with GenBuster-200K real videos), (d) FaceForensics++ (C23, original vs. all manipulated classes pooled), (e) CelebDF [ 4 ] (cf. Section V-E classification results).