Generalized Design Choices for Deepfake Detectors

Lorenzo Pellegrini; Serafino Pandolfini; Davide Maltoni; Matteo Ferrara; Marco Prati; Marco Ramilli

cs.CVNov 26, 2025

Generalized Design Choices for Deepfake Detectors

Authors: Lorenzo Pellegrini, Serafino Pandolfini, Davide Maltoni, Matteo Ferrara, Marco Prati, Marco Ramilli

Organizations: Department of Computer Science and Engineering (DISI) University of Bologna, Italy · IdentifAI, Italy

Abstract

The effectiveness of deepfake detection methods often depends less on their core design and more on implementation details such as data preprocessing, augmentation strategies, and optimization techniques. These factors make it difficult to fairly compare detectors and to understand which factors truly contribute to their performance. To address this, we systematically investigate how different design choices influence the accuracy and generalization capabilities of deepfake detection models, focusing on aspects related to training, inference, and incremental updates. By isolating the impact of individual factors, we aim to establish robust, architecture-agnostic best practices for the design and development of future deepfake detection systems. Our experiments identify a set of design choices that consistently improve deepfake detection and enable state-of-the-art performance on the AI-GenBench benchmark.

Figures & tables

Figure 1 : The different dimensions explored in this work, to optimize the training, inference and incremental update (grayed) of deepfake detectors.

Figure 2 : Augmentation pipeline. Error bars denote the standard deviation across the three runs.

Pipeline	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0	Average
Baseline	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06	90.11
Evaluation	95.63 $\pm$ 0.42	94.76 $\pm$ 0.19	92.77 $\pm$ 0.43	90.22 $\pm$ 0.12	93.34
Mild	95.44 $\pm$ 0.40	94.24 $\pm$ 0.62	89.95 $\pm$ 0.72	89.71 $\pm$ 0.08	92.33

Table 1: Effect of the training-time augmentation pipeline on the Next Period AUROC averaged over all temporal evaluation steps (%). Values are mean

\pm

standard deviation over the three runs; the Average column reports the mean across the different backbones. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 3 : Augmentation multiplier. Error bars denote the standard deviation across the three runs.

$am$	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0	Average
1	87.93 $\pm$ 0.69	82.45 $\pm$ 2.18	75.94 $\pm$ 0.69	80.61 $\pm$ 4.01	81.73
2	91.71 $\pm$ 0.33	88.19 $\pm$ 1.54	81.07 $\pm$ 1.00	86.32 $\pm$ 0.16	86.82
3	93.95 $\pm$ 0.32	90.80 $\pm$ 0.97	83.08 $\pm$ 0.81	87.84 $\pm$ 0.07	88.92
4	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06	90.11
5	95.33 $\pm$ 0.39	92.89 $\pm$ 0.93	86.49 $\pm$ 1.25	89.42 $\pm$ 0.07	91.03
6	95.72 $\pm$ 0.42	93.40 $\pm$ 0.99	87.56 $\pm$ 0.42	89.88 $\pm$ 0.04	91.64

Table 2: Effect of the augmentation multiplier (

am

) on the Next Period AUROC averaged over all temporal evaluation steps (%). Values are mean

\pm

standard deviation over the three runs; the Average column reports the mean across backbones. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 4 : Training duration. Error bars denote the standard deviation across the three runs.

Epochs	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0	Average
1	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06	90.11
2	96.47 $\pm$ 0.17	94.18 $\pm$ 0.68	89.07 $\pm$ 0.67	90.50 $\pm$ 0.12	92.55
3	96.96 $\pm$ 0.12	94.32 $\pm$ 0.74	90.67 $\pm$ 0.52	91.30 $\pm$ 0.09	93.31
4	97.05 $\pm$ 0.13	94.19 $\pm$ 0.76	91.82 $\pm$ 0.23	91.81 $\pm$ 0.08	93.72

Table 3: Effect of the training duration (epochs) on the Next Period AUROC averaged over all temporal evaluation steps (%). Values are mean

\pm

Figure 5 : Input processing. Error bars denote the standard deviation across the three runs.

Training input	Evaluation input	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Resize	Resize	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06
Resize	Mixed	94.75 $\pm$ 0.31	91.77 $\pm$ 1.15	82.18 $\pm$ 0.88	86.67 $\pm$ 0.10
Crop	Mixed	94.73 $\pm$ 0.31	88.42 $\pm$ 0.82	84.56 $\pm$ 1.59	86.50 $\pm$ 0.08

Table 4: Comparison of training and evaluation input strategies (Next Period AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

standard deviation over the three runs. Mixed refers to the fusion of the scores obtained from the crops and from the resized version of the full image. The first row corresponds to the Baseline model. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 6 : Plain multiclass training. Error bars denote the standard deviation across the three runs.

Fusion strategy	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06
Sum fusion	92.47 $\pm$ 0.37	85.65 $\pm$ 1.38	81.79 $\pm$ 1.64	87.78 $\pm$ 0.10
Max fusion	91.84 $\pm$ 0.44	85.35 $\pm$ 1.33	81.77 $\pm$ 1.44	88.54 $\pm$ 0.11

Table 5: Plain multiclass training: comparison of fusion strategies (Next Period AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

standard deviation over the three runs. Baseline refers to direct binary training. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 7 : Dual-head training. Error bars denote the standard deviation across the three runs.

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06
Separate heads	92.84 $\pm$ 0.33	90.20 $\pm$ 1.01	81.89 $\pm$ 1.05	88.27 $\pm$ 0.21
Stacked heads	91.52 $\pm$ 0.30	81.47 $\pm$ 7.02	78.00 $\pm$ 1.35	85.96 $\pm$ 1.10
Separate heads (aux)	94.02 $\pm$ 0.46	92.53 $\pm$ 0.56	82.32 $\pm$ 1.57	88.67 $\pm$ 0.10
Stacked heads (aux)	93.11 $\pm$ 0.48	85.46 $\pm$ 2.76	79.36 $\pm$ 2.05	85.64 $\pm$ 0.80

Table 6: Dual-head training results (Next Period AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

standard deviation over the three runs. Aux denotes down-weighting the multiclass loss (

w_{bin}=0.75,w_{multi}=0.25

). Baseline refers to training using a binary loss only. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 8 : Distance-based scoring. Error bars denote the standard deviation across the three runs.

Centroids	Fusion	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	–	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06
1	Sum	85.69 $\pm$ 1.29	84.55 $\pm$ 1.09	71.10 $\pm$ 0.72	79.42 $\pm$ 0.31
2	Sum	86.01 $\pm$ 0.32	85.59 $\pm$ 1.22	71.46 $\pm$ 1.47	78.73 $\pm$ 0.13
3	Sum	85.95 $\pm$ 0.86	85.97 $\pm$ 0.91	71.04 $\pm$ 1.23	77.76 $\pm$ 0.54
1	Max	83.75 $\pm$ 0.66	84.20 $\pm$ 1.20	75.74 $\pm$ 3.32	75.95 $\pm$ 0.21
2	Max	80.45 $\pm$ 0.48	84.98 $\pm$ 1.22	70.43 $\pm$ 2.82	76.07 $\pm$ 0.32

Table 7: Distance-based scoring vs. MLP baseline (Next Period AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

standard deviation over the three runs. The best result per backbone (highest mean) is highlighted in bold. Bootstrap 95% confidence intervals are reported in Appendix A .

Figure 9 : Continual learning, Next Period. Error bars denote the standard deviation across the three runs.

Figure 10 : Continual learning, Past Period. Error bars denote the standard deviation across the three runs.

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline (reference)	94.56 $\pm$ 0.45	92.05 $\pm$ 1.16	85.08 $\pm$ 0.52	88.76 $\pm$ 0.06
Reset weights	91.16 $\pm$ 0.18	88.63 $\pm$ 1.69	80.14 $\pm$ 1.33	86.95 $\pm$ 0.13
Naive	91.68 $\pm$ 0.27	89.18 $\pm$ 0.72	80.28 $\pm$ 0.14	87.06 $\pm$ 0.15
Replay (CB, 10k)	92.37 $\pm$ 0.29	92.00 $\pm$ 1.17	81.12 $\pm$ 1.56	87.18 $\pm$ 0.19
Replay (CB, 20k)	92.82 $\pm$ 0.43	92.44 $\pm$ 1.35	81.31 $\pm$ 0.81	87.10 $\pm$ 0.19
Replay (harmonic)	93.37 $\pm$ 0.49	92.89 $\pm$ 0.95	83.36 $\pm$ 0.67	87.20 $\pm$ 0.19

Table 8: Next Period performance of different training regimes (AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

standard deviation over the three runs. The best continual learning result per backbone (highest mean) is highlighted in bold. CB stands for class-balanced , followed by the number of samples in the replay buffer. Bootstrap 95% confidence intervals are reported in Appendix A .

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline (reference)	99.27 $\pm$ 0.09	98.20 $\pm$ 0.34	91.97 $\pm$ 0.56	95.36 $\pm$ 0.08
Reset weights	97.80 $\pm$ 0.19	96.17 $\pm$ 0.51	85.78 $\pm$ 1.62	93.34 $\pm$ 0.09
Naive	96.19 $\pm$ 0.52	93.72 $\pm$ 0.23	82.12 $\pm$ 0.77	91.83 $\pm$ 0.11
Replay (CB, 10k)	98.14 $\pm$ 0.06	97.37 $\pm$ 0.55	86.79 $\pm$ 1.78	93.07 $\pm$ 0.06
Replay (CB, 20k)	98.57 $\pm$ 0.24	97.73 $\pm$ 0.68	87.47 $\pm$ 0.65	93.35 $\pm$ 0.07
Replay (harmonic)	98.71 $\pm$ 0.16	98.03 $\pm$ 0.49	89.59 $\pm$ 0.66	93.32 $\pm$ 0.10

Table 9: Past Period performance of different training regimes (AUROC averaged over all temporal evaluation steps, %). Values are mean

\pm

Window	Baseline	Naive	CB (10k)	CB (20k)	Harmonic
1	100%	50.00%	65.63%	75.00%	75.00%
2	100%	33.33%	43.75%	54.17%	58.33%
3	100%	25.00%	32.81%	40.63%	47.92%
4	100%	20.00%	26.25%	32.50%	40.83%
5	100%	16.67%	21.88%	27.08%	35.69%
6	100%	14.29%	18.75%	23.21%	31.79%

Table 10: Relative computational cost per window as a percentage of the baseline. The setup uses AI-GenBench (32k new examples/window, divided into 16k real and 16k fake), with window 24 simulating a 20-year benchmark. CB stands for class-balanced, followed by the number of samples in the replay buffer.

Recommendation	Backbone	AUROC $\Delta$ (pp)	Train	Infer.
Evaluation augmentation vs. baseline (Fig. 2 )	All	$+3.23$ avg.	$1\times$	$1\times$
Four epochs vs. one epoch (Fig. 4 )	All	$+3.61$ avg.	$4\times$	$1\times$
Mixed inference vs. resize-only (Fig. 5 )	DINOv2	$+0.19$	$1\times$	$6\times$
	ViT-L CLIP	$-0.27$
	ResNet-50 CLIP	$-2.90$
	EfficientNet-B0	$-2.09$

Table 11: Accuracy–cost summary for the main recommendations. AUROC differences are expressed in percentage points (pp) and are computed from the three-run means reported in the previous tables. Mixed inference uses one resized input plus five crops and is applied to resize-trained models. For replay, the Next Period average is reported. The corresponding 95% bootstrap confidence intervals are reported in Table 21 .

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Pipeline	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
Evaluation	95.68 [95.53, 95.83]	94.55 [94.37, 94.74]	93.22 [93.03, 93.41]	90.19 [89.96, 90.41]
Mild	95.43 [95.27, 95.59]	93.53 [93.33, 93.72]	90.22 [90.00, 90.45]	89.69 [89.45, 89.92]

Appendix

Table 12: Training-time augmentation pipelines. Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

$am$	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
1	87.46 [87.20, 87.73]	79.93 [79.61, 80.26]	75.92 [75.56, 76.29]	75.98 [75.62, 76.33]
2	91.73 [91.52, 91.95]	86.44 [86.16, 86.72]	81.77 [81.45, 82.09]	86.51 [86.24, 86.78]
3	93.86 [93.68, 94.04]	89.71 [89.45, 89.96]	82.65 [82.33, 82.96]	87.91 [87.65, 88.17]
4	94.22 [94.04, 94.40]	90.79 [90.55, 91.02]	85.21 [84.92, 85.51]	88.79 [88.54, 89.03]
5	95.14 [94.99, 95.31]	91.83 [91.60, 92.06]	86.54 [86.26, 86.82]	89.40 [89.16, 89.63]
6	95.73 [95.57, 95.88]	92.30 [92.07, 92.52]	87.46 [87.19, 87.73]	89.87 [89.64, 90.11]

Appendix

Table 13: Augmentation multiplier (

am

), with the number of training epochs fixed to one. Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Epochs	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
1	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
2	96.38 [96.25, 96.52]	93.41 [93.20, 93.61]	89.20 [88.95, 89.44]	90.39 [90.16, 90.61]
3	96.94 [96.81, 97.07]	93.46 [93.26, 93.67]	90.68 [90.45, 90.91]	91.21 [91.00, 91.43]
4	97.09 [96.97, 97.21]	93.31 [93.10, 93.53]	92.04 [91.82, 92.25]	91.73 [91.52, 91.93]

Appendix

Table 14: Training duration (epochs), with

am=4

. Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Train. / Eval. input	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Resize / Resize	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
Resize / Mixed	94.63 [94.45, 94.80]	90.65 [90.41, 90.90]	82.81 [82.49, 83.13]	86.79 [86.52, 87.06]
Crop / Mixed	94.95 [94.78, 95.11]	87.48 [87.21, 87.76]	82.72 [82.41, 83.02]	86.54 [86.27, 86.81]

Appendix

Table 15: Training and evaluation input strategies; rows follow the order of Table 4 (resize/resize, resize/mixed, crop/mixed). Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Fusion strategy	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
Sum fusion	92.87 [92.67, 93.07]	84.55 [84.27, 84.82]	82.81 [82.50, 83.12]	87.86 [87.59, 88.11]
Max fusion	92.31 [92.10, 92.52]	84.55 [84.26, 84.83]	82.64 [82.33, 82.95]	88.67 [88.41, 88.91]

Appendix

Table 16: Plain multiclass training with sum and max fusion; rows follow the order of Table 5 . Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
Separate heads	93.15 [92.96, 93.35]	89.03 [88.79, 89.27]	82.99 [82.68, 83.30]	88.05 [87.79, 88.30]
Stacked heads	91.17 [90.95, 91.39]	84.95 [84.69, 85.20]	77.70 [77.40, 78.01]	86.08 [85.81, 86.35]
Separate heads (aux)	94.21 [94.03, 94.39]	92.35 [92.14, 92.55]	84.09 [83.79, 84.39]	88.55 [88.30, 88.80]
Stacked heads (aux)	93.66 [93.47, 93.85]	86.74 [86.49, 87.00]	81.65 [81.34, 81.96]	85.65 [85.37, 85.92]

Appendix

Table 17: Dual-head training; rows follow the order of Table 6 . Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Centroids	Fusion	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline	–	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
1	Sum	86.69 [86.41, 86.96]	84.14 [83.81, 84.46]	71.84 [71.44, 72.23]	79.43 [79.09, 79.76]
2	Sum	86.35 [86.07, 86.63]	85.23 [84.92, 85.54]	70.88 [70.48, 71.27]	78.63 [78.28, 78.97]
3	Sum	86.73 [86.45, 87.00]	84.93 [84.62, 85.24]	70.48 [70.08, 70.86]	77.58 [77.23, 77.93]
1	Max	83.15 [82.86, 83.43]	83.98 [83.65, 84.31]	77.71 [77.33, 78.07]	76.10 [75.75, 76.45]
2	Max	80.88 [80.58, 81.18]	84.63 [84.31, 84.95]	72.10 [71.70, 72.49]	76.17 [75.83, 76.52]

Appendix

Table 18: Distance-based scoring; the first three data rows use sum fusion and the last three max fusion, as in Table 7 . Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline (reference)	94.22 [94.04, 94.40]	90.79 [90.55, 91.03]	85.21 [84.93, 85.51]	88.79 [88.54, 89.03]
Reset weights	91.01 [90.79, 91.22]	86.69 [86.43, 86.96]	81.08 [80.75, 81.41]	87.09 [86.83, 87.36]
Naive	91.83 [91.62, 92.03]	88.36 [88.10, 88.62]	80.15 [79.81, 80.48]	87.17 [86.91, 87.43]
Replay (CB, 10k)	92.14 [91.94, 92.34]	90.66 [90.43, 90.90]	82.19 [81.88, 82.51]	87.31 [87.05, 87.57]
Replay (CB, 20k)	93.29 [93.10, 93.48]	90.96 [90.72, 91.19]	81.66 [81.33, 81.99]	87.23 [86.98, 87.49]
Replay (harmonic)	93.76 [93.57, 93.94]	91.87 [91.65, 92.09]	84.13 [83.82, 84.43]	87.03 [86.77, 87.30]

Appendix

Table 19: Incremental update strategies, Next Period metric; rows follow the order of Table 8 . Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Next Period AUROC averaged over all temporal evaluation steps (%).

Configuration	DINOv2	ViT-L CLIP	ResNet-50 CLIP	EfficientNet-B0
Baseline (reference)	99.24 [99.20, 99.28]	97.81 [97.74, 97.89]	92.41 [92.27, 92.55]	95.42 [95.33, 95.51]
Reset weights	97.98 [97.93, 98.03]	95.79 [95.70, 95.88]	86.65 [86.49, 86.81]	93.41 [93.31, 93.51]
Naive	96.50 [96.45, 96.56]	93.62 [93.53, 93.72]	82.13 [81.95, 82.30]	91.92 [91.82, 92.03]
Replay (CB, 10k)	98.20 [98.16, 98.25]	96.74 [96.65, 96.82]	88.03 [87.88, 88.19]	93.12 [93.02, 93.22]
Replay (CB, 20k)	98.83 [98.79, 98.86]	96.95 [96.87, 97.03]	87.40 [87.24, 87.56]	93.43 [93.33, 93.53]
Replay (harmonic)	98.88 [98.85, 98.92]	97.46 [97.38, 97.54]	90.36 [90.21, 90.51]	93.20 [93.10, 93.31]

Appendix

Table 20: Incremental update strategies, Past Period metric; rows follow the order of Table 9 . Entries are the single-run point estimate followed by the 95% bootstrap confidence interval, for the Past Period AUROC averaged over all temporal evaluation steps (%).

Recommendation	Backbone	AUROC $\Delta$ (pp) [95% CI]
Evaluation augmentation vs. baseline	All	$+3.66$ avg. $[+3.23,+4.09]$
Four epochs vs. one epoch	All	$+3.79$ avg. $[+3.36,+4.22]$
Mixed inference vs. resize-only	DINOv2	$+0.41$ $[+0.06,+0.76]$
	ViT-L CLIP	$-0.13$ $[-0.62,+0.35]$
	ResNet-50 CLIP	$-2.40$ $[-3.02,-1.80]$
	EfficientNet-B0	$-2.00$ $[-2.52,-1.48]$

Appendix

Table 21: 95% bootstrap confidence intervals of the AUROC differences summarized in Table 11 . Differences and intervals are computed from the single-run point estimates of the compared configurations and are reported here for completeness; Table 11 reports the same differences computed from the three-run means.

Explore similar work

Deepfake Detection Generalization with Diffusion Noise
Apr 16, 2026Hongyuan Qi, Wenjin Hou, Hehe Fan +1Deepfake Detection Unsupervised Detection
Zhejiang University Hangzhou, Zhejiang, China
$μ$ Flow: Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors
Jun 29, 2026Orazio Pontorno, Mattia Litrico, Luca Guarnera +2Deepfake Detection Unsupervised Detection
University of Catania, Catania, Italy · University of Nottingham, Nottingham, UK
Continuously Evolving Deepfake Detection: An Architecture and Public-Benchmark Evaluation of a Dynamic Detection System
Jul 14, 2026Ken Jon Miyachi, Dylan UysDeepfake Detection Digital Forensics
BitMind

Apr 16, 2026, cs.CVJ/K move · Enter open · S save

Deepfake Detection Generalization with Diffusion Noise

Hongyuan Qi, Wenjin Hou, Hehe Fan, Jun Xiao

Zhejiang University Hangzhou, Zhejiang, China

Deepfake Detection Unsupervised Detection Noise-Aware

Open paper AskPDF arXiv

Deepfake detectors face growing challenges in generalization as new image synthesis techniques emerge. In particular, deepfakes generated by diffusion models are highly photorealistic and often evade detectors trained on GAN-based forgeries. This paper addresses the generalization problem in deepfake detection by leveraging diffusion noise characteristics. We propose an Attention-guided Noise Learning (ANL) framework that integrates a pre-trained diffusion model into the deepfake detection pipeline to guide the learning of more robust features. Specifically, our method uses the diffusion model's denoising process to expose subtle artifacts: the detector is trained to predict the noise contained in an input image at a given diffusion step, forcing it to capture discrepancies between real and synthetic images, while an attention-guided mechanism derived from the predicted noise is introduced to encourage the model to focus on globally distributed discrepancies rather than local patterns. By harnessing the frozen diffusion model's learned distribution of natural images, the ANL method acts as a form of regularization, improving the detector's generalization to unseen forgery types. Extensive experiments demonstrate that ANL significantly outperforms existing methods on multiple benchmarks, achieving state-of-the-art accuracy in detecting diffusion-generated deepfakes. Notably, the proposed framework boosts generalization performance (e.g., improving ACC/AP by a substantial margin on unseen models) without introducing additional overhead during inference. Our results highlight that diffusion noise provides a powerful signal for generalizable deepfake detection.

Abstract

Explore similar work

Deepfake Detection Generalization with Diffusion Noise

μμμFlow: Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors

Continuously Evolving Deepfake Detection: An Architecture and Public-Benchmark Evaluation of a Dynamic Detection System

$μ$ Flow: Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors