DeepFense: A Unified, Modular, and Extensible Framework for Robust Deepfake Audio Detection

Yassine El Kheir; Arnab Das; Yixuan Xiao; Xin Wang; Feidi Kallel; Enes Erdem Erdogan; Ngoc Thang Vu; Tim Polzehl

cs.SDApr 9, 2026

DeepFense: A Unified, Modular, and Extensible Framework for Robust Deepfake Audio Detection

Authors: Yassine El Kheir, Arnab Das, Yixuan Xiao, Xin Wang, Feidi Kallel, Enes Erdem Erdogan, Ngoc Thang Vu, Tim Polzehl, +1 more

Organizations: German Research Center for Artificial Intelligence (DFKI), Germany · University of Stuttgart, Germany · National Institute of Informatics, Japan · Technical University of Berlin, Germany

Abstract

Speech deepfake detection is a well-established research field with different models, datasets, and training strategies. However, the lack of standardized implementations and evaluation protocols limits reproducibility, benchmarking, and comparison across studies. In this work, we present DeepFense, a comprehensive, open-source PyTorch toolkit integrating the latest architectures, loss functions, and augmentation pipelines, alongside over 100 recipes. Using DeepFense, we conducted a large-scale evaluation of more than 400 models. Our findings reveal that while carefully curated training data improves cross-domain generalization, the choice of pre-trained front-end feature extractor dominates overall performance variance. Crucially, we show severe biases in high-performing models regarding audio quality, speaker gender, and language. DeepFense is expected to facilitate real-world deployment with the necessary tools to address equitable training data selection and front-end fine-tuning.

Figures & tables

Figure 1: System architecture of the DeepFense framework, illustrating the configuration-driven data pipeline, modular model engine (front-end, back-end, loss), unified training loop, and evaluation components, with built-in logging and experiment tracking.

	ITW		ASV5		CodecFake		Average
Back-End	Orig.	Ours	Orig.	Ours	Orig.	Ours	Orig.	Ours
AASIST	10.46	7.18	19.24	18.94	38.79	34.36	22.83	20.16
MLP	8.13	7.57	18.76	15.43	35.76	34.44	20.88	19.15
Nes2Net	7.06	7.73	22.06	24.19	39.34	36.21	22.82	22.71
TCM	7.79	6.98	18.85	19.40	36.00	34.77	20.88	20.39

Table 1: EER (%) for Wav2Vec2 systems trained on ASV19 , evaluated on out-of-domain test sets before and after applying DeepFense. Best result per back-end in bold .

Back-End	Orig.	Ours
AASIST	7.12	6.64
Nes2Net	6.13	5.53
BiCrossMamba-ST	6.01	6.34

Table 2: EER (%) for Wav2Vec2 systems trained on ASV5 , evaluated on the ASV5 test set.

Back-End	Orig.	Ours
Nes2Net	0.49	0.55
AASIST	1.03	0.77
MLP	0.95	0.91
TCM	0.34	0.40
BiCrossMamba-ST	0.75	0.44

Table 3: EER (%) for EAT trained on EnvSDD and evaluated on CodecFake-A3 test set.

Figure 2: Boxplots of detection systems based on their EERs on the test set (x-axis). Each box groups 16 systems using the same front-end (from left to right: Wav2Vec2, WavLM, HuBERT, and EAT) but varied training data and back-ends. The three lines correspond to mean EERs of systems using either the ASV19 (blue), ADD23 (red), or CodecFake (green) training sets.

Figure 3: Mean EER (%) per front-end and evaluation benchmark, averaged over all back-ends and training sets.

Front-End	Mean EER (%)	Back-End	Mean EER (%)
Wav2Vec2	25.5	AASIST	30.3
EAT	31.4	Nes2Net	30.7
WavLM	32.4	TCM	30.8
Hubert	33.6	MLP	31.1

Table 4: Mean EER (%) for each front-end (averaged over all back-ends & training sets) and each back-end (averaged over all front-ends & training sets).

\downarrow

Panel of Figure 4 — Figure 4: Mean EER analysis across training datasets, front-ends, and back-ends. (a) Front-end perspective. (b) Back-end perspective.

Training Data	Mean EER	Std
CodecFake	22.3	0.047
HABLA	25.2	0.034
ASV5	26.9	0.029
ASV19	28.2	0.052
PartialSpoof	31.0	0.056
ADD23	50.8	0.063

Table 5: Macro-average EER across all 13 evaluation datasets, per training set (mean

\pm

std over all front-end

\times

back-end).

Figure 5: EER on non-speech deepfake detection.

Figure 6: EER Sensitivity to Audio Quality Across Systems trained on ASV19, and CodecFake

			GARBE
Training	Front-End	Back-End	PESQ	NISQA-MOS	Gender	Language
ASV19	EAT	AASIST	0.0135	0.0196	0.1460	0.1886
	Hubert	AASIST	0.1735	0.1626	0.1160	0.2397
	Wav2Vec2	AASIST	0.1880	0.2396	0.1940	0.2147
	Wav2Vec2	TCM	0.1769	0.2056	0.2110	0.2630
	WavLM	AASIST	0.2579	0.2258	0.1450	0.2841

Table 6: GARBE fairness score of each system, computed over audio quality (PESQ and NISQA-MOS), gender, and languages.

Training	Front-End	Back-End	F	M	$\Delta$
			EER
ASV19	EAT	AASIST	0.584	0.578	$+0.005$
	Hubert	AASIST	0.447	0.321	$+0.126$
	Wav2Vec2	AASIST	0.257	0.183	$+0.075$
	Wav2Vec2	TCM	0.275	0.200	$+0.074$
	WavLM	AASIST	0.455	0.347	$+0.108$

Table 7: EER by gender for each system. F = Female, M = Male,

\Delta=\text{EER}_{\text{F}}-\text{EER}_{\text{M}}

Figure 7: EER (%) per language for all system configurations trained on ASVspoof 2019 (left) and CodecFake (right). Languages are sorted by mean EER (lowest at top). Each dot is one system; colour encodes front-end; marker shape encodes back-end. Grey bars show the min–max EER range across all configurations.

Explore similar work

What You Train Is What You Get: Gender Bias, Training Composition, and Post-Hoc Mitigation in Audio Deepfake Detection
Jul 10, 2026Aishwarya R. Fursule, Vamshi Nallaguntla, Shruti Kshirsagar +1Audio Deepfake Detection ASVspoof
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
Teffic-Audio: Tell Fact from Fiction
Jul 30, 2026Wan Lin, Li Wang, Jindong Wang +2Audio Deepfake Detection Domain Generalization
Amphion Team
Probing-Guided Layer Selection from Self-Supervised Speech Models for Generalizable Audio Deepfake Detection
Jun 29, 2026Marjan Beheshti, Majid Rostami, Bo ChenAudio Deepfake Detection Domain Generalization
Department of Computer Science, Michigan Technological University, Houghton, MI, USA

Jul 10, 2026J/K move · Enter open · S save

What You Train Is What You Get: Gender Bias, Training Composition, and Post-Hoc Mitigation in Audio Deepfake Detection

Aishwarya R. Fursule, Vamshi Nallaguntla, Shruti Kshirsagar, Anderson R. Avila

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

Audio Deepfake Detection ASVspoof

Open paper AskPDF arXiv

Audio deepfake detection models determine whether speech is genuine or artificially generated, but high overall accuracy can mask substantial performance disparities across demographic groups. In this work, we investigate gender bias in audio deepfake detection using the ASVspoof5 dataset. We use ASVspoof5 under a controlled custom split designed to isolate gender-composition effects. We train attack-specific models on nine training sets with different gender compositions, ranging from female-only to male-only. We use a ResNet18 classifier with LogSpectrogram and WavLM-Base+ features, and we evaluated six post-hoc threshold calibration methods. Experimental results show that training data composition strongly predicts bias direction, with the underrepresented gender performing worse at test time. WavLM-Base+ features are shown to produce gender performance gaps 3.0 to 4.3 times larger than LogSpectrogram under identical training conditions, and balanced training is found to reduce LogSpectrogram bias but leave WavLM bias largely intact. Moreover, all six calibration strategies, including Oracle calibration with full test-set label access, leave the Equal Error Rate gap unchanged at 1.317 pp, confirming that threshold adjustment cannot correct underlying score distribution disparities. Overall, these findings suggest that gender fairness in audio deepfake detection must be addressed at training time, as post-hoc methods can only partially mitigate the resulting disparities

Abstract

Explore similar work

What You Train Is What You Get: Gender Bias, Training Composition, and Post-Hoc Mitigation in Audio Deepfake Detection

Teffic-Audio: Tell Fact from Fiction

Probing-Guided Layer Selection from Self-Supervised Speech Models for Generalizable Audio Deepfake Detection