acceptodds
Under review as a conference paper at ICLR 2027

GLAD: Global-Local Adaptive Detector for Robust Speech Deepfake Detection

Abstract

Recent advances in AI-based speech synthesis have enabled highly realistic speech, increasing the importance of speech deepfake detection (SDD) in preventing misuse. While mainstream Self-Supervised Learning (SSL)-based detectors achieve strong performance, they suffer from poor generalization to unseen domains and often overlook fine-grained signal artifacts due to a bias towards global semantic consistency. In this paper, we conduct the first detailed empirical and visual analysis to validate these limitations explicitly. Our investigation reveals two critical architectural vulnerabilities: (1) a systemic failure to capture localized spoofing traces, and (2) a severe lack of adaptability to domain-driven shifts in SSL layer importance, rendering static aggregation strategies prone to overfitting. To address these vulnerabilities, we propose the Global-Local Adaptive Detector (GLAD). Specifically, to capture localized forgeries, GLAD employs a Hierarchical Global-Local (HGL) backbone that explicitly bridges the granularity gap by fusing global linguistic and acoustic features with fine-grained local signal details. To counter layer importance shifts in out-of-distribution (OOD) scenarios, we introduce a Hierarchical Adaptive Gating (HAG) mechanism that dynamically recalibrates layer-wise focus in a sample-specific manner. Finally, to address shortcut learning induced by environmental biases, we introduce SaniBoost, a composite data augmentation strategy for robust signal standardization and noise sanitization. Extensive experiments demonstrate that GLAD significantly outperforms state-of-the-art methods, particularly on unseen domain cases. Codes are available at the anonymous repository.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Speech Deepfake Detection, Cross-Domain Robustness, Anti Spoofing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WkVuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1(−2)
Soundness
2
Read review
Paper summary

This paper proposes GLAD, a Global-Local Adaptive Detector for robust speech deepfake detection. The framework integrates three feature extraction branches, including XLS-R 300M, WavLM-Large, and a signal-level CNN, through a Hierarchical Global-Local (HGL) backbone. It further introduces Hierarchical Adaptive Gating (HAG) to dynamically adjust layer-wise and stream-wise contributions, together with SaniBoost, a data augmentation strategy designed to mitigate environmental shortcut learning. The model is trained on ASVspoof 2019 LA and evaluated across several in-domain and out-of-domain benchmarks, including ASVspoof 2021 LA/DF, PartialSpoof, In-the-Wild, and Fake-or-Real.

Critical strengths
  1. The paper addresses the important problem of generalization in speech deepfake detection, particularly the limitations of SSL-based representations under unseen synthesis methods and recording conditions.

  2. The proposed framework provides a systematic integration of complementary SSL representations and signal-level features. The combination of global-local modeling, adaptive layer aggregation, and data augmentation is practically meaningful.

  3. The experimental evaluation covers multiple benchmarks and demonstrates strong cross-domain performance. The paper also includes multi-seed evaluations, component ablations, capacity-controlled comparisons, and additional analyses of local spoofing attribution, providing useful empirical evidence for the overall framework.

Critical weaknesses
  1. The fairness of the SOTA comparison requires further clarification.
    Although GLAD is trained on ASVspoof 2019 LA and evaluated across five benchmarks, it is unclear whether the competing methods in Table 1 follow the same training and augmentation protocols. In particular, representative baselines such as XLS-R & SLS should be evaluated under the same SaniBoost strategy to better isolate the architectural contribution. Appendix F.2 provides useful controlled comparisons, but these do not fully address the fairness of the main SOTA evaluation.

  2. The individual contributions of the three backbones are insufficiently evaluated.
    GLAD combines XLS-R, WavLM-Large, and a raw-waveform CNN. However, the "w/o HGL" ablation simultaneously removes multiple components. Comparisons such as XLS-R + WavLM, XLS-R + CNN, and WavLM + CNN are missing. In particular, the XLS-R + WavLM configuration is important for determining whether the local CNN branch provides substantial complementary information beyond the two SSL encoders.

  3. The contribution of SaniBoost is not fully isolated.
    Although Table 14 compares RawBoost and several ABFP variants, it remains unclear whether the improvements arise from bona fide purification itself or from reducing RawBoost augmentation on bona fide samples. A probability-matched control replacing ABFP with no augmentation would help clarify this distinction. The individual effects of RMS normalization, phase inversion, and Gaussian noise also require further analysis.

  4. The oracle layer importance analysis lacks sufficient methodological clarity.
    Figure 2 uses layer weights learned from target-domain training as oracle references. However, the construction of these training sets for ASVspoof 2021 DF and In-the-Wild, which lack official training partitions, is not clearly described. Potential training/evaluation overlap and the validity of these weights as oracle references should be clarified. Furthermore, DTW permits cross-layer alignment despite fixed SSL layer positions. Position-wise metrics would provide more direct evidence, although cosine similarity is additionally reported in the appendix.

  5. The semantic-acoustic specialization of the SSL encoders is not sufficiently justified.
    The paper designates XLS-R as the semantic encoder and WavLM as the acoustic encoder, although both can encode linguistic and acoustic information. Direct probing analysis or experiments exchanging their roles in the fusion architecture would help verify whether this functional distinction is empirically supported.

Questions for authors
  1. The fairness of the SOTA comparison requires further clarification.
    Although GLAD is trained on ASVspoof 2019 LA and evaluated across five benchmarks, it is unclear whether the competing methods in Table 1 follow the same training and augmentation protocols. In particular, representative baselines such as XLS-R & SLS should be evaluated under the same SaniBoost strategy to better isolate the architectural contribution. Appendix F.2 provides useful controlled comparisons, but these do not fully address the fairness of the main SOTA evaluation.

  2. The individual contributions of the three backbones are insufficiently evaluated.
    GLAD combines XLS-R, WavLM-Large, and a raw-waveform CNN. However, the "w/o HGL" ablation simultaneously removes multiple components. Comparisons such as XLS-R + WavLM, XLS-R + CNN, and WavLM + CNN are missing. In particular, the XLS-R + WavLM configuration is important for determining whether the local CNN branch provides substantial complementary information beyond the two SSL encoders.

  3. The contribution of SaniBoost is not fully isolated.
    Although Table 14 compares RawBoost and several ABFP variants, it remains unclear whether the improvements arise from bona fide purification itself or from reducing RawBoost augmentation on bona fide samples. A probability-matched control replacing ABFP with no augmentation would help clarify this distinction. The individual effects of RMS normalization, phase inversion, and Gaussian noise also require further analysis.

  4. The oracle layer importance analysis lacks sufficient methodological clarity.
    Figure 2 uses layer weights learned from target-domain training as oracle references. However, the construction of these training sets for ASVspoof 2021 DF and In-the-Wild, which lack official training partitions, is not clearly described. Potential training/evaluation overlap and the validity of these weights as oracle references should be clarified. Furthermore, DTW permits cross-layer alignment despite fixed SSL layer positions. Position-wise metrics would provide more direct evidence, although cosine similarity is additionally reported in the appendix.

  5. The semantic-acoustic specialization of the SSL encoders is not sufficiently justified.
    The paper designates XLS-R as the semantic encoder and WavLM as the acoustic encoder, although both can encode linguistic and acoustic information. Direct probing analysis or experiments exchanging their roles in the fusion architecture would help verify whether this functional distinction is empirically supported.

1 comment

Thank you for your detailed responses and the substantial additional experiments. I particularly appreciate the authors' positive and constructive attitude toward addressing the reviewers' concerns. The newly provided controlled comparisons, backbone ablations, and augmentation analyses have satisfactorily addressed most of my questions and strengthened the experimental evidence. In light of these improvements, I am willing to increase my rating from 2 to 3.

One minor point remains regarding Q4. Although the rebuttal clarifies the training setup and the construction of the target-domain datasets used for the oracle analysis, I could not locate these details in the manuscript itself. I encourage the authors to incorporate this information into the paper for clarity and reproducibility.

Additionally, I suggest highlighting the newly added experiments and corresponding revisions in the manuscript so that other reviewers can easily locate and assess these improvements.

Overall, I appreciate the authors' efforts and the additional evidence provided during the rebuttal.

Reviewer_yemcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets cross-domain generalization in speech deepfake detection. GLAD fuses two self-supervised speech encoders (XLS-R and WavLM) with a raw-waveform CNN branch through hierarchical cross-attention and a global-local fusion stage. A Hierarchical Adaptive Gating (HAG) module computes, for each utterance, gates over encoder layers, over the two streams, and over the local branch. SaniBoost, the training-time augmentation, normalizes loudness and phase and applies aggressive high-pass filtering and noise gating to a subset of bona fide utterances (Adversarial Bona Fide Purification, ABFP), so that background noise cannot serve as a cue for authenticity. The model is trained on ASVspoof 2019 LA and evaluated on ASVspoof 2021 LA (21LA), ASVspoof 2021 DF (21DF), PartialSpoof, In-the-Wild (ITW) and Fake-or-Real. The appendix adds three-seed results, capacity-matched controls that share the same backbones, an attribution analysis on short partial spoofs, and an intervention that replaces the learned gates.

Critical strengths
  1. The full system's gains survive controls for seeds, backbones and parameter count. Table 3 reports mean ± std over three seeds. Table 17 (Appendix F.2) gives five alternatives the same two encoders, the same CNN branch, the same augmentation and the same development-set selection rule. Against the parameter-matched MLP (662.24M vs 662.23M parameters), GLAD lowers mean EER from 4.87% to 2.70% on 21LA, from 2.74% to 1.53% on 21DF, and from 6.63% to 5.91% on ITW. This rules out the simplest explanation, that a second backbone or more parameters produce the gains.

  2. The functional claims come with testable numbers. Table 6 (Appendix C) shows better attribution alignment on short partial spoofs on all three metrics in every segment-count group. Table 7 (Appendix E.1) takes the same trained checkpoint and replaces each utterance's gates with the target-domain mean gate, estimated on the other half of the data (cross-fitting). EER rises on all three datasets, and the paired-bootstrap 95% intervals for the increase exclude zero (21LA +0.4255 [0.3268, 0.5223], 21DF +0.1601 [0.1277, 0.2048], ITW +0.1777 [0.0592, 0.2877]). This supports a real role for per-utterance gating and for the local branch.

Critical weaknesses
  1. The gain over the closest existing method is not isolated. Section 3.4 describes the layer scorer as refining "the standard SLS strategy by replacing the single linear scoring layer with a Multi-Layer Perceptron", and Section 2.1 acknowledges that the original Sensitive Layer Selection (SLS) already computes input-conditioned layer weights. Yet no experiment swaps only the scorer:

    (a) The "XLS-R & SLS" baseline in Table 1 uses a single XLS-R backbone and a different fusion, so its gap to GLAD mixes the scorer with everything else.

    (b) The "static SLS" control in Table 17 uses sample-independent layer weights with late fusion, which is not what the original SLS does.

    The results therefore show that the full GLAD system beats these alternatives, not how much the proposed scorer adds over the original adaptive SLS inside the same architecture. Since that refinement is part of the claimed contribution, this missing comparison bears directly on the assessment.

  2. Several comparison numbers are wrong or cannot be traced.

    (a) AASIST row in Table 4. The row lists A18/A19/EER/min t-DCF as 1.52/3.40/0.62/1.13. In the original AASIST paper (Table 1, https://arxiv.org/pdf/2110.01200) the mean values are A17 1.52, A18 3.40, A19 0.62, pooled EER 1.13 and min t-DCF 0.0347. A min t-DCF of 1.13 is outside the range of that metric and is in fact AASIST's pooled EER, so the row has been shifted by one column. The entries 0.16 (A11) and 0.15 (A15) also do not appear in the source.

    (b) Yang et al. (2024) row in Table 5. Table 5 attributes 21LA = 6.56 and no ITW result to this paper. The cited version (https://arxiv.org/pdf/2403.01960) evaluates only on ASVspoof 2019 LA, ASVspoof 2021 DF and ITW, reports 24.27% EER on ITW for its fusion system, and contains no 21LA result and no value of 6.56. The "940M" parameter label is also unexplained: that paper uses HuBERT-base and WavLM-Base-Plus alongside XLS-R.

    (c) GLAD's own A18/A19 values. Table 4 gives 0.57/0.31; Appendix B.4 gives means of 0.46/0.30 and then 0.43/0.28, without saying which seeds or checkpoints produced each pair.

    These errors do not touch the matched controls in Table 17, but they lower confidence in every comparison table until the sources and per-seed results are given.

Questions for authors
  1. Scorer comparison. Is the "static SLS" in Table 17 intentionally different from the original input-conditioned SLS? Keeping the encoders, fusion, augmentation and selection rule fixed, what happens when only the HAG layer scorer is replaced by the original SLS scorer? This result would directly change my assessment of the contribution.

  2. Comparison tables.

    • (a) Please correct the AASIST row in Table 4.
    • (b) Please give the exact source, or the rerun setup, for the Yang et al. row in Table 5, including the parameter count.
    • (c) Please state the seeds and checkpoints behind the three A18/A19 pairs in Table 4 and Appendix B.4.
    • (d) After these corrections, which comparative conclusions change?
  3. SaniBoost mechanism. The paper says ABFP removes "environmental shortcut cues". Appendix E.2 shows that a model trained with ABFP is stable under the ABFP transformation itself. Which analysis distinguishes "reduced reliance on environmental cues" from "robustness to the specific transformation used in training"?

Reviewer_xrQKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue of insufficient generalization of self-supervised speech forgery detection models in out-of-distribution scenarios. It proposes a unified framework, GLAD, which combines global-local feature fusion (HGL), sample-adaptive hierarchical gating (HAG), and data augmentation (SaniBoost). The detection performance advantages are verified through multiple cross-domain benchmarks and controlled ablation experiments.

Critical strengths

Strengths:

S1. It has highly competitive cross-domain detection performance and comprehensive validation coverage. GLAD achieves competitive results on multiple evaluation sets when trained using only ASVspoof 2019 LA.

S2. The ablation experiments are relatively complete and can support the effectiveness of the overall architecture design. The paper not only provides the overall ablation of HGL, HAG, and SaniBoost, but also further breaks down cross-attention, feature fusion, layer aggregation, flow gating, and data augmentation strategies in Appendix Table 14.

Critical weaknesses

Main argument:

  1. The causal explanations in the core research motivation remain stronger than the conclusions that experiments can prove. I could summarize the core design idea in two parts: First, the cross-domain failure of the SSL detector primarily stems from insufficient capture of local forgery traces; second, the changing distribution of SSL layer importance between the source and target domains is a significant reason for the model's inadequate generalization. However, while existing experiments can demonstrate a correlation between these phenomena, they are insufficient to establish them as the root cause of cross-domain failures. Figure 1 and Appendix Table 6 show that GLAD's attribution results are closer to the labeled local fake regions than XLS-R-SLS. This supports the improvement in local sensitivity, but it cannot yet directly prove that local branches of CNN are the key source of cross-domain performance improvement. In particular, the overall Frame AUPRC of GLAD in Table 6 is 0.2811 and IoU@GT is 0.1565, indicating that the improvement clearly exists, but the absolute level of localization capability is still limited.

  2. Regarding the changes in layer importance, Figure 2 uses the model retrained in the target domain to generate a so-called oracle distribution. A core reasoning problem exists here: the layer preferences formed after training in the target domain are not necessarily equivalent to the optimal layer preferences that the source domain model should adopt when testing in the target domain.

  3. According to Appendix A.2, SaniBoost performs special cleaning on 35% of the bona fide samples, while RawBoost is used for the remaining bona fide samples and all spoof samples. This enhancement strategy may indeed improve generalization performance, but it also introduces an alternative explanation that needs to be ruled out: the model may learn special processing traces associated with the labels, rather than reducing its reliance on environmental cues. Appendix Table 9 shows that this strategy produces fairly strong acoustic transformations, including a median energy drop of approximately 40 dB below 500 Hz, and noise gating that zeros out approximately 77% of the waveform samples at the median.

  4. The bidirectional cross-attention mechanism of HGL, the local feature injection of CNNs, the Sigmoid gating and MLP layer weighting of HAG, and the various signal transformations of SaniBoost are all closely related to existing deep learning methods. The real contribution of this paper lies in how to organize these mechanisms for speech forgery detection and the cross-domain performance obtained after combination, rather than proposing entirely new fundamental operations or learning paradigms.

Questions for authors
  1. How much of GLAD's performance improvement comes from the proposed architecture rather than the increased number of parameters and pretrained encoders? Could the authors provide comparisons with parameter-matched models using the same backbones?

  2. Are the improvements over strong SSL-based baselines consistent, particularly on 21DF and In-the-Wild?

  3. What evidence demonstrates that HAG performs genuinely sample-specific adaptation rather than learning a fixed preference for each domain or class? Would replacing the dynamic gates with averaged gates affect performance?

  4. Does GLAD's improved sensitivity to local spoofing artifacts generalize beyond the illustrative example in Figure 1? Could the authors provide quantitative evidence on PartialSpoof, especially for very short or fragmented spoofed regions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.