The rapid advancement of generative models has enabled highly realistic audio deepfakes, yet current detectors suffer from a critical bias problem, leading to poor generalization across unseen datasets. This paper proposes Artifact-Focused Self-Synthesis (AFSS), a method designed to mitigate this bias by generating pseudo-fake samples from real audio via two mechanisms: self-conversion and self-reconstruction. The core insight of AFSS lies in enforcing same-speaker constraints, ensuring that real and pseudo-fake samples share identical speaker identity and semantic content. This forces the detector to focus exclusively on generation artifacts rather than irrelevant confounding factors. Furthermore, we introduce a learnable reweighting loss to dynamically emphasize synthetic samples during training. Extensive experiments across 7 datasets demonstrate that AFSS achieves state-of-the-art performance with an average EER of 5.45%, including a significant reduction to 1.23% on WaveFake and 2.70% on In-the-Wild, all while eliminating the dependency on pre-collected fake datasets. Our code is publicly available at https://github.com/NguyenLeHaiSonGit/AFSS.
Figures & tables
Fig. 1: Cross-dataset performance of audio deepfake detection methods (EER%) when trained on ASV19-LA and tested on different datasets. Lower values indicate better performance.
Fig. 2: Overview of the proposed AFSS method.
Fig. 3: Overview of model architecture.
ASVspoof 19
ASVspoof 21
Real-World
Method
Venue
eval
DF eval
DF hidden
LA eval
LA hidden
In-the-Wild
WaveFake
Avg
AASIST-B [ 8 ]
ICASSP ’22
1.10
16.70
21.16
7.40
32.81
39.45
45.45
23.44
AASIST-L [ 8 ]
ICASSP ’22
1.58
18.75
20.11
8.57
26.47
32.61
43.95
21.72
SCL [ 4 ]
Interspeech ’24
2.88
2.17
5.76
13.90
15.15
4.52
2.99
6.77
Conformer [ 18 ]
Interspeech ’23
0.44
4.84
8.37
1.67
12.19
8.52
18.82
7.84
SSL [ 30 ]
ICASSP ’24
1.91
5.67
8.84
15.92
14.97
6.10
1.30
7.82
TABLE I: Cross-dataset generalization performance comparison (EER%). Lower is better. Best results are in bold .
ASVspoof 19
ASVspoof 21
Real-World
Method
Venue
eval
DF eval
DF hidden
LA eval
LA hidden
In-the-Wild
WaveFake
Avg
AASIST-B [ 8 ]
ICASSP ’22
0.9984
0.9059
0.8389
0.9811
0.7229
0.6531
0.9165
0.8595
AASIST-L [ 8 ]
ICASSP ’22
0.9982
0.8989
0.8652
0.9702
0.7985
0.7372
0.9250
0.8847
SCL [ 4 ]
Interspeech ’24
0.9984
0.9968
0.9851
0.9343
0.9200
0.9910
0.9901
0.9737
Conformer [ 18 ]
Interspeech ’23
0.9995
0.9903
0.9720
0.9942
0.9418
0.9712
0.8868
0.9651
AFSS (Ours)
0.9939
0.9965
0.9794
0.9610
0.9440
0.9964
0.9990
0.9815
TABLE II: Cross-dataset generalization performance comparison (AUC). Higher is better. Best results are in bold .
ASVspoof 19
ASVspoof 21
Real-World
Method
Venue
LA eval
DF eval
DF hidden
LA eval
LA hidden
In-the-Wild
WaveFake
ALDEN [ 33 ]
ACM MM ’25
1.22
4.39
-
-
-
11.27
4.66
DSVAE [ 34 ]
2.16
-
-
-
-
-
-
SafeEar [ 12 ]
CCS ’24
3.10
-
-
7.22
-
-
-
AFSS (Ours)
2.72
2.19
6.92
10.02
12.35
2.70
1.23
TABLE III: Comparison with disentanglement methods (EER%). Results are taken from original publications; ’-’ indicates the metric was not reported by the original authors. Best results are in bold .
Components
ASV19-eval
In-the-Wild
EER
AUC
ACC
AP
EER
AUC
ACC
AP
S-VC + S-Rec
8.57
97.81
91.44
99.75
4.08
99.14
95.92
98.57
+ Alpha
6.61
97.37
93.39
99.71
3.19
99.52
96.81
99.19
+ Bal
6.84
97.31
93.16
99.69
3.74
99.29
96.26
98.81
+ Alpha + Bal
2.72
99.39
97.26
99.93
2.70
99.64
97.30
99.39
TABLE IV: Effectiveness of the proposed components (%). S-VC: Self-conversion, S-Rec: Self-reconstruction, Alpha: Learnable Reweighting Loss, Bal: Balanced Batch Sampling.
Dataset
Intensity 0
Intensity 1
Intensity 2
Intensity 3
ASV19 eval
1.71%
2.72%
9.01%
7.45%
DF21 eval
2.65%
2.19%
3.15%
5.74%
DF21 hidden
9.04%
6.92%
11.82%
13.64%
LA21 eval
8.63%
10.02%
13.96%
16.23%
LA21 hidden
12.55%
12.35%
15.05%
15.97%
WaveFake
6.46%
1.23%
14.10%
5.63%
TABLE V: Effects of data transformation intensity (EER%)
LR
Alpha value
EER
ACC
AUC
AP
10−4
wf=1.73,wr=0.63
1.63%
98.37%
99.81%
99.98%
10−5
wf=1.77,wr=0.68
2.97%
97.02%
99.31%
99.93%
10−6
wf=1.77,wr=0.69
1.23%
98.77%
99.90%
99.99%
TABLE VI: Optimization stability analysis: learning rate effects on reweighting parameters. Results are on WaveFake dataset.
Method
ASV19-eval
ASV21-DF
WaveFake
In-the-Wild
Cross-Speaker VC
5.68
21.25
33.11
55.98
AFSS (Ours)
2.72
2.19
1.23
2.70
TABLE VII: Comparison of Equal Error Rate (EER, %) between our proposed AFSS and a standard cross-speaker voice conversion (VC) baseline.
School of Computing, Wichita State University, Wichita, KS, USA · Institut national de la recherche scientifique (INRS–EMT), Montreal, QC, Canada · INRS-UQO Mixed Research Unit on Cybersecurity, Gatineau, QC, Canada
School of Computing, Wichita State University, Wichita, KS, USA · Institut national de la recherche scientifique (INRS-EMT), Montreal, QC, Canada · INRS-UQO Mixed Research Unit on Cybersecurity, Gatineau, Canada