While deepfake speech detectors built on large self-supervised learning (SSL) models achieve high accuracy, employing standard ensemble fusion to further enhance robustness often results in oversized systems with diminishing returns. To address this, we propose an evolutionary multi-objective score fusion framework that jointly minimizes detection error and system complexity. We explore two encodings optimized by NSGA-II: binary-coded detector selection for score averaging and a real-valued scheme that optimizes detector weights for a weighted sum. Experiments on the ASVspoof 5 dataset with 36 SSL-based detectors show that the obtained Pareto fronts outperform simple averaging and logistic regression baselines. The real-valued variant achieves 2.37% EER (0.0684 minDCF) and identifies configurations that match state-of-the-art performance while significantly reducing system complexity, requiring only half the parameters. Our method also provides a diverse set of trade-off solutions, enabling deployment choices that balance accuracy and computational cost.
Figures & tables
Fig. 1 : Overview of the proposed evolutionary multi-objective fusion framework. The system processes input recordings through a pool of base detectors to obtain scores and parameter counts. A candidate fusion is constructed and evaluated based on a randomly initialized chromosome. NSGA-II then iteratively optimizes the fusion configuration (i.e., the chromosome) to simultaneously minimize the Equal Error Rate (EER) and system complexity (i.e., the number of parameters), ultimately producing a Pareto front of optimal trade-off solutions.
Fig. 2 : Chromosome representations for the two variants.
Fig. 3 : Impact of population size on HV convergence under fixed computational budget (best viewed in color). Averages of 10 runs are reported.
Parameter
Binary Variant
Real-Valued Variant
Encoding
c∈{0,1}D
c∈[0,1]D
Population Size ( N )
100
100
Selection
Binary Tournament
Binary Tournament
Crossover Type
Uniform
Uniform
Crossover Rate ( pc )
0.7
0.5
Mutation Type
Bit-flip
Polynomial
TABLE II : Hyperparameters of the NSGA-II Fusions
Fig. 4 : HV convergence – averages of 10 runs are reported.
Fig. 5 : Parameter sensitivity analysis for the real-valued variant: mean HV vs. pm and pc for three distinct distribution indices ( ηm ) under a fixed computational budget of 25,000 fitness evaluations.
Fig. 6 : Parameter sensitivity analysis for the binary variant: mean HV vs. pm and pc under fixed computational budget.
Fig. 7 : Multi-objective deepfake speech detector fusions comparing EER [%] (horizontal axis) against number of parameters (vertical axis). The red and purple markers represent the global non-dominated solutions (super-Pareto front) found by 10 runs of the NSGA-II real-valued and binary variants, respectively. Single points for logistic regression (blue markers) and manual averaging (orange markers) are shown for comparison, representing baseline performance and complexity. Best viewed in color.
System
EER
minDCF
Parameters
T27 [ 16 ]
3.42%
0.0937
approx. 286M
T23 [ 31 ]
3.41%
0.0936
approx. 475M
T43 (Winner) [ 5 ]
2.59%
0.0750
approx. 5B
Ours
4.36%
0.1268
416M
Ours
3.38%
0.0965
720M
Ours
2.59%
0.0747
2.49B
TABLE III : Comparison with ASVspoof 5 top teams. Selected real-valued NSGA-II solutions provide direct reference points to the teams in terms of either EER or parameter count.
Speech deepfake detectors usually emit one score per utterance, but a borderline score does not reveal why two examples differ during retrospective error analysis. We ask whether a final score can be calibrated from component fields while keeping those fields visible for inspection. We build a decision record with a passive detector score and a score from a probe applied to a marked copy. It also includes retrieval support held out of the evaluated family, a margin from a support-set profile, and raw neighbor closeness. A cross-fit calibrator combines these fields and two differences between raw scores into one final score. On matched ASVspoof development data, the calibrated record reduces equal error rate (EER) by 3.48 percentage points relative to the fixed retrieval-augmented rule. It reaches 8.43% EER, whereas a passive WavLM baseline reaches 6.71% on the same subset. The record is therefore not the strongest detector in this comparison. Its value is to retain inspectable component fields while producing one scalar score for retrospective diagnosis.
Mengzhe Geng, Yujia Lu, Manuela Kunz +1
National Research Council Canada · The Chinese University of Hong Kong
This paper describes our submission to the ICME 2026 ESDD2 challenge on environment-aware speech and sound deepfake detection. The task requires five-class classification of audio clips in which speech, environmental sound, both components, or neither component may be spoofed. We propose a component-level ensemble system based on four publicly available pre-trained anti-spoofing models: XLSR-Mamba, DF-Arena, SLS, and TCM-ADD. Each model is fine-tuned on the official CompSpoofV2 development data using three binary heads for original, speech, and environmental sound detection. We further train RawBoost-augmented variants and combine selected checkpoints using margin-space score fusion. A component-wise fusion strategy with lightweight head- and class-bias calibration yields our best configuration, reaching 0.7715 macro-F1 on the evaluation set and 0.7828 macro-F1 on the test set, ranking 5th out of 31 teams in the final ranking phase and substantially outperforming the official baseline.
André Runewicz, Karla Schäfer, Martin Steinebach
Fraunhofer SIT — ATHENE – National Research Center for Applied Cybersecurity Darmstadt, Germany
This paper describes a submission to the Environment-Aware Speech and Sound Deepfake Detection Challenge (ESDD2) 2026, which addresses component-level deepfake detection using the CompSpoofV2 dataset, where speech and environmental sounds may be independently manipulated. To address this challenge, a dual-branch deepfake detection framework is proposed to jointly model speech and environmental contextual representations from input audio. Two pretrained models, XLS-R for speech and BEATs for environmental sound, are used to extract complementary contextual representations. A Matching Head is introduced to model representation differences through statistical normalization and representation interaction, enabling estimation of the original class. In parallel, multi-head cross-attention enables effective information exchange between speech and environmental components. The refined representations are processed with residual connections and layer normalization, and passed to an AASIST classifier to predict speech-based and environment-based spoofing probabilities. The model outputs original, speech, and environment predictions. On the test set, the proposed system achieves an F1-score of 70.20% and an environmental EER of 16.54%, outperforming the baseline system.
Khalid Zaman, Qixuan Huang, Muhammad Uzair +1
Graduate School of Advanced Science and Technology · Japan Advanced Institute of Science and Technology · 1-1 Asahidai, Nomi, Ishikawa, Japan 923-1292