Explainable Audio Deepfake Detection

• Explainable audio deepfake detection (ADD) is a field that distinguishes genuine from manipulated audio while providing transparent, localized rationales.
• Methodologies such as diffusion-based segmentation, transformer attributions, occlusion visualizations, and chain-of-thought reasoning yield robust, interpretable outputs.
• Enhanced explainability in ADD systems improves forensic analysis by identifying manipulation regions, attributing sources, and mitigating adversarial vulnerabilities.

Explainable audio deepfake detection (ADD) is the research field focused on not only distinguishing bona fide from manipulated or AI-synthesized audio, but also generating accurate, interpretable rationales for model decisions. True explainability in ADD systems entails the ability to localize manipulated regions in time and frequency, attribute the type and origin of manipulations, quantify evidence supporting the decision, and provide outputs that are robust and informative under adversarial conditions. Requirements for explainable ADDs extend beyond traditional binary classifiers: they must provide transparent and localized insight into the detection process, enabling forensics, building trust, and improving real-world reliability.

1. Problem Formulation and Importance of Explainability

Explainable ADD extends standard binary detection by requiring:

• Temporal and spectral localization: Identification of specific regions in the time–frequency space altered by manipulation (“when”).
• Manipulation traceability: Determination of the nature of the manipulation or generative method (“how” and “where”).
• Source attribution: Fingerprinting the generative or editing source of a forged segment.

Explainability is pivotal for forensics and legal evidence, as practitioners must know what was changed and by which tool. Trustworthy systems expose their rationale for classification so human analysts can verify and potentially contest results. Further, explainability constrains models to rely on semantically meaningful features, mitigating overfitting to dataset-specific artifacts or spurious cues (Xie et al., 21 Sep 2025).

2. Methodological Approaches to Explainable ADD

Several families of methodologies are employed in explainable ADD:

a) Diffusion-based Artifact Localization

A prominent approach involves treating artifact localization as a supervised segmentation problem. Given parallel bona fide and fake (e.g., vocoded) audio, a ground-truth mask is computed in the time–frequency domain by taking the magnitude difference of their respective short-time Fourier transforms (STFTs), smoothing and normalizing to suppress noise, then thresholding to extract artifact regions (Grinberg et al., 3 Jun 2025):

• For $x_{\mathrm{real}}(t)$ and $x_{\mathrm{voc}}(t)$, $$\mathrm{STFT}(x_{\mathrm{real}}) = M_b(t,f) + i P_b(t,f)$$, $$\mathrm{STFT}(x_{\mathrm{voc}}) = M_s(t,f) + i P_s(t,f)$$.
• Raw difference: $$D_{\mathrm{raw}}(t, f) = \left| M_s(t, f) - M_b(t, f) \right|$$.
• Mask: $H^*(t, f) = 1$ if $$| \overline{M}_s(t, f) - \overline{M}_b(t, f) | / | \overline{M}_b(t, f) | > \tau$$, otherwise 0.

A conditional diffusion model (SegDiff-style DDPM with U-Net) is then trained to predict artifact segmentation masks given the fake audio’s spectrogram or deep features, using standard denoising objectives.

b) Transformer-Relevant and Gradient-Based Attribution

For transformer-based models, transformer-aware relevancy maps (e.g., GATR) propagate attention and gradient information layer by layer, assigning a continuous importance score $R(t)$ to each timestep. This reveals which temporal (or spectrotemporal) portions drive the model’s decisions (Grinberg et al., 23 Jan 2025). Baselines include Grad-CAM (adapted to 1-D or time–frequency), DeepSHAP, and GradientSHAP.

c) Occlusion- and Attention-Rollout Visualizations

Occlusion saliency involves masking segments or patches of the spectrogram to measure their influence on classification probability, producing heatmaps localizing critical artifacts. Attention rollout compounds layer-wise attention matrices to yield effective input token importance—interpretable as proxies for salient regions in the signal (Channing et al., 2024).

d) Audio LLMs with Frequency–Time Chain-of-Thought (CoT) Reasoning

Large audio LLMs (ALLMs) or audio LLMs receive audio tokens (via embedding aligners) and prompts, producing outputs structured as chain-of-thought rationales that explicitly reference spectral and temporal anomalies. Reinforcement fine-tuning with custom frequency–time reward functions constrains the model to generate evidence-traceable explanations, enhancing both detection accuracy and human verifiability (Xie et al., 6 Jan 2026).

3. Quantitative and Qualitative Evaluation Protocols

Explainability techniques are evaluated using a multifaceted protocol:

Metric Type	Examples and Descriptions
Segmentation/Local.	GDice, F1, IoU: Alignment with ground-truth artifact mask
Faithfulness	AI (Avg. Increase), AD (Avg. Drop), AG (Avg. Gain), Fid-In (Input Fidelity)
Attribution	Acc_manipulation (manipulation type), Acc_source (generator fingerprinting)
Localization	F1_segment (temporal), Temporal IoU
Robustness	Perturbation (AUC-EER), OOD evaluation (e.g., on unseen generators/sources)
Reasoning Audit	Perception, coherence, dissonance metrics (for model reasoning reliability)

Segmentation and faithfulness metrics are computed by perturbing audio according to explanation masks and measuring effects on the detector’s output, or by quantifying overlap of predicted importance with known forged regions (Grinberg et al., 3 Jun 2025, Grinberg et al., 23 Jan 2025). Attribution is quantified via accuracy in identifying manipulation type/source (Xie et al., 21 Sep 2025).

Explainability audits on LLM-based models assess cognitive coherence (support of rationale for decision), perception score (match with ground-truth forensic questions), and cognitive dissonance (useful alarm for silent misclassification) (Nguyen et al., 7 Jan 2026).

Qualitative evaluation involves overlaying masks, heatmaps, or rationales on spectrogram displays for expert inspection (Channing et al., 2024).

4. Benchmarks and Comparative Experimental Findings

The introduction of large-scale, multi-task benchmarks such as FakeSound2 has standardized evaluation along three axes: (1) localization of manipulated regions, (2) traceability (type and source), and (3) generalization to out-of-domain manipulations. Experiments reveal:

• In-domain: Strong localization (F1_segment ~97%) and source attribution (Acc_source ~85%) (Xie et al., 21 Sep 2025).
• Out-of-domain: Large performance drops in traceability—Acc_manipulation from 72% to 42%, sometimes approaching zero for particular generation types—while binary detection remains relatively robust.

Comparative results for explanation techniques demonstrate that data-driven, diffusion-based segmentation models (ADDSegDiff, SpecSegDiff) significantly outperform classical post-hoc XAI methods (DeepSHAP, GradientSHAP, AttnLRP) both in mask alignment and faithfulness (Grinberg et al., 3 Jun 2025):

Method	GDice/F1 (Dev/Test)	AI/AD (Dev/Test)
DeepSHAP	11/9, 6.5/4.2	—
AttnLRP	13/9, 8.3/4.2	—
ADDSegDiff	48/52, 45.7/49.6	0.81/1.82, 68/85
SpecSegDiff	57/52, 55.3/49.3	—

On partial-spoof datasets, transformer-aware relevancy explanations (GATR) yield the most faithful and localized explanations, outperforming Grad-CAM and SHAP-based attributions as measured by Relevance Rank Accuracy and Relevance Mass Accuracy (Grinberg et al., 23 Jan 2025).

5. Audio LLMs, Chain-of-Thought Reasoning, and Robustness

ALLMs fine-tuned with frequency–time chain-of-thought rationales (FT-GRPO) achieve state-of-the-art detection across all audio types (speech, sound, singing, music), with co-trained AVG accuracy of 90.1% (Xie et al., 6 Jan 2026). Under FT-GRPO, generated rationales are explicitly grounded in spectral bands and temporal regions, making forensic verification tractable. Supervised cold-start followed by policy optimization under FT constraints prevents rationale hallucination, remedying deficiencies of both SFT (black-box output) and vanilla RL (reward hacking).

Forensic audits under adversarial conditions reveal a bifurcation (Nguyen et al., 7 Jan 2026):

• Models with strong acoustic grounding (e.g., Qwen2, granite) gain robustness (“shield effect”) from explicit reasoning, reducing attack success rates in linguistic adversarial settings.
• Models with weaker perceptual fidelity suffer degradation (“reasoning tax”), with increased adversarial vulnerability due to hallucinated or incoherent rationales.

Cognitive dissonance metrics enable “silent alarms” on misclassifications, especially under acoustic attacks, by flagging cases where a model’s reasoning contradicts its verdict.

6. Challenges, Limitations, and Research Directions

Significant open challenges and limitations remain:

• Ground-truth dependencies: Diffusion-based artifact localization requires parallel bona fide/fake pairs and may not generalize to edit-based or non-vocoded manipulations (Grinberg et al., 3 Jun 2025).
• Source fingerprinting: Models often overfit to dataset-specific artifacts rather than intrinsic generator characteristics, hobbling traceability and generalization (Xie et al., 21 Sep 2025).
• Limited rationalization under adversarial shifts: Reasoning chains can be undermined by acoustic or linguistic attacks, sometimes masking or even exacerbating failures (Nguyen et al., 7 Jan 2026).
• Method variance on large-scale tests: XAI explanations derived from small samples do not always generalize; broad dataset-level evaluation is essential (Grinberg et al., 23 Jan 2025).

Future research recommendations include:

• Multi-task training for joint localization, type/source attribution, and detection (Xie et al., 21 Sep 2025).
• Contrastive fingerprinting losses and domain-adapted augmentation for OOD robustness.
• Incorporation of phoneme/prosody conditioning and interpretable module design for richer rationale production (Grinberg et al., 3 Jun 2025, Xie et al., 6 Jan 2026).
• Cognitive dissonance monitoring and adversarial training to stabilize explanation reliability in high-stakes applications (Nguyen et al., 7 Jan 2026).
• Development of quantitative explanation quality metrics, counterfactual reasoning modules, and expanded, multilingual benchmarks (Channing et al., 2024, Xie et al., 21 Sep 2025).

A plausible implication is that future practical ADD systems will blend robust interpretable front-ends (e.g., attention- or prototype-based transformers, diffusion-based artifact localizers) with structured, LLM-driven rationale generation, integrated with user-facing visualization and alarm metrics.

7. Practical System Integration and Forensic Workflows

State-of-the-art explainable ADD pipelines combine detection and artifact localization as follows (Grinberg et al., 3 Jun 2025):

1. Waveform-based ADD model $f$ generates a real/fake score.
2. If spoofed, the system extracts a log-magnitude spectrogram or intermediate backbone features.
3. A diffusion-based segmentation model (e.g., SpecSegDiff, ADDSegDiff) generates a heatmap highlighting artifact regions.
4. Explanations are overlaid on spectrograms for expert review.

For LLM-based rationale explainers, outputs consist of (a) natural-language frequency/time artifact lists, (b) explicit localization tags, and (c) final verdicts, optionally accompanied by reasoning disagreement alarms.

These practices aim to equip forensic analysts, legal authorities, and end-users with inspection tools capable of both highly accurate detection and transparent, human-verifiable outputs, marking a transition toward accountable and trustworthy audio deepfake detection frameworks.