acceptodds
Under review as a conference paper at ICLR 2027

Learning as Deepfakes Evolve: RF-Prompt for Continual Audio Deepfake Detection

Abstract

Speech generation methods are evolving rapidly, creating a moving target for audio deepfake detection (ADD). A deployed detector must incorporate newly emerging deepfake methods without forgetting previously learned real and deepfake knowledge. Continual learning provides a natural solution, but existing continual ADD evaluations commonly define tasks by dataset, coupling changes in real-speech sources with changes in deepfake mechanisms and obscuring what knowledge is being updated. We address task organization and detector adaptation jointly. First, we construct five protocols from identical training, development, and evaluation pools. Among them, the proposed real-anchored mechanism-incremental (RAMI) protocol reflects a practical detector-update scenario: real speech from known source domains is available, while newly arriving deepfake mechanisms must be learned continually without forgetting earlier knowledge. Second, we propose RF-Prompt (real–fake prompt learning), an asymmetric continual prompt-learning method. A shared real prompt provides protected adaptation capacity, while task-specific fake experts expand as new mechanisms arrive. Parameter-level cosine anchoring stabilizes the real prompt; each new fake expert inherits a selected historical expert and learns a residual with soft orthogonal regularization. Input-adaptive fusion integrates the accumulated fake experts without requiring task identity at inference. RF-Prompt obtains its lowest common-average and pooled EER under RAMI among the five tested protocols. On RAMI, it achieves 10.110% final average EER and 10.370% pooled EER, the lowest aggregate errors among the evaluated continual-learning baselines. Ablations assess the adaptation components, while limited-training-data and cross-backbone experiments examine their applicability across settings.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Audio deepfake detection, Continual learning, Prompt learning, Mixture of experts

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_DcSROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies continual learning for audio deepfake detection, aiming to learn emerging deepfake methods without forgetting previously acquired knowledge. It makes two main contributions:

RAMI Evaluation Protocol: Organizes incremental tasks by four generation mechanisms while maintaining a consistent mixture of real speech sources, addressing the confounding of real speech domains and fake generation methods in dataset-based protocols. Five protocols are constructed for comparison. RF-Prompt: Combines a shared Real Prompt with expanding Fake Prompt Experts. It uses cosine regularization to preserve shared knowledge, expert inheritance and residual orthogonality to learn new knowledge, and Soft MoE to adaptively fuse experts. The authors evaluate RF-Prompt against several continual learning baselines, with additional ablation, limited-data, and cross-backbone experiments. Results show improved detection performance and reduced forgetting, and ablations support the effectiveness of the key components.

Critical strengths

Meaningful and practical problem setting: The paper addresses an important challenge in audio deepfake detection: adapting to emerging generation methods while retaining previously learned knowledge. The proposed RAMI protocol provides a more targeted evaluation setting than conventional dataset-based task organizations.

Well-motivated and coherent method design: RF-Prompt combines shared knowledge preservation with expandable fake experts. Historical expert inheritance and residual orthogonality encourage knowledge transfer and complementary learning, while adaptive fusion enables inference without task identity.

Comprehensive experimental evaluation: The paper compares against a range of continual learning baselines and includes ablation studies, limited-data experiments, and cross-backbone evaluations. The results demonstrate competitive detection performance and reduced forgetting, with ablations supporting the effectiveness of key components.

Critical weaknesses

Confounding between generation mechanisms and dataset sources: Although RAMI aims to isolate different generation mechanisms, the training samples for M1, M3, and M4 each come from a single dataset. This makes it difficult to distinguish mechanism-specific knowledge from dataset-specific characteristics.

Limited evaluation of long-term continual learning: RAMI groups generation methods into only four broad mechanisms, making it unclear how newly emerging generators within existing mechanisms can be incorporated as additional tasks. Moreover, all experiments use a single random seed, leaving the stability of the reported improvements uncertain.

Insufficient validation of key design choices: While the ablation studies support the effectiveness of the main components, some design choices lack direct evidence. For example, both Real and Fake Prompts are trained on both classes without explicit class-specific constraints, making it unclear whether the Real Prompt specifically preserves real-speech knowledge.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer DcSR

Thank you for recognizing the practical motivation and coherent design. Following your suggestions, we added the experiments below.

W1. Mechanism/source confounding

We added separate budget and source controls. Primary epochs are complete passes; the new experiments explicitly match optimization settings.

Matched-budget control: 50 epochs/task, 200 updates/epoch, and 16 real plus 16 fake utterances/batch. All organizations are evaluated on the same 40,000 utterances (Table 4, seed 2026).

fake organizationCommon Avg EER (%)Pool EER (%)
Dataset-wise (P4)11.44511.830
Random four groups11.65811.972
Mechanism-wise (RAMI)10.42210.738

Same-source control: identical AT-ADD M2/M4 subsets and real samples; 50 epochs/task, 75 updates/epoch, and 16+16 utterances/batch for both streams (Table 16, seed 2026).

fake organizationCommon Avg EER (%)Pool EER (%)
Random two groups5.7505.450
M2 then M45.3205.125

The first control matches budget and batch balance; the second holds dataset identity fixed. Together they support mechanism organization. Appendix I gives settings, and Appendix A documents the full mechanism/source composition.

W2. New generators within a mechanism and stability

M1-M4 are families, not a maximum of four updates. Following your suggestion, we added generator-incremental streams using all original samples and 40 training generators. The 10-task order is M1a/M2a/M3a/M4a/M1b/M2b/M3b/M4b/M2c/M3c; the 20-task stream splits each batch into two.

Seed 2026; Table 5:

TasksRAWM Pool EER (%)RF-Prompt Pool EER (%)
1012.50510.595
2012.56010.910

These experiments directly test later generators from recurring mechanism families.

We also added two seeds per method. Table 1 now reports seeds 2026/2027/2028. RF-Prompt achieves Avg EER 10.118 +/- 0.058% and Pool EER 10.402 +/- 0.076% (mean +/- sample SD), the lowest mean aggregate EER among continual methods. Supplementary diagnostics remain seed 2026.

W3. Why associate the shared prompt with real knowledge?

The sharing choice follows the practical update scenario. We have real speech from known domains and finite observed deepfake data, while new methods continually emerge. A protected shared prompt reuses real reference knowledge; fake experts expand for new patterns.

Related motivation comes from ASDG (real aggregation), OC-Softmax (compact real representations), and LSR/LSA (one real/multiple fake prototypes), already cited in Section 2.

Both prompt types receive binary supervision and participate in every prediction. Their names denote intended roles; real-reference knowledge need not be learned with real-only classification updates.

Added structure/plasticity controls: Table 12 favors shared-real/expanding-fake over its reverse and symmetric alternatives. Freezing the shared branch is also worse than protected adaptation. The table below includes the original real-anchoring ablation as a reference (seed 2026):

Shared-branch settingPool EER (%)
No cosine anchoring (Table 3)11.365
Freeze after Task 1 (Appendix K.3)12.675
Protected adaptation (RF-Prompt)10.370

Added class-conditioned sensitivity experiment: after zeroing the shared prompt, the mean absolute change in predicted real probability is:

Input classRange across M1-M4
real0.362-0.376
fake0.016-0.081

The branch has a stronger influence on real-example scores. These are absolute probability changes on a 0-1 scale, not EER improvements (Table 21). The training controls establish detection benefits; Appendix K.2 characterizes score sensitivity.

Reviewer_fYKZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on continual learning for audio deepfake detection.

First, it introduces the RAMI (Real-Anchored Mechanism-Incremental) protocol, in which fake audio samples are learned sequentially according to their generation mechanisms, rather than following the conventional dataset-wise task formulation.

Second, it proposes RF-Prompt, which combines a shared real prompt with task-specific fake experts.

Under the RAMI protocol, RF-Prompt demonstrates the best overall performance among the compared methods.

Critical strengths
  • [S1] Practical and Realistic Problem Setting

The setting in which a detector is updated whenever a new deepfake generation technique emerges is both realistic and important.

  • [S2] Controlled Protocol Comparison

The use of the same training, development, and evaluation pools across five protocols is a strength of the study. In particular, the comparison between Protocol 4 and RAMI allows the authors to investigate the impact of fake-task organization while keeping the real-speech provisioning strategy fixed.

  • [S3] Well-Motivated Asymmetric Architecture

The design of preserving a shared real prompt while progressively expanding task-specific fake experts is well motivated. In particular, the decomposition $ P_{f,t}=B_t+\Delta_t $ is interesting because it explicitly distinguishes between inheriting previously acquired knowledge and incorporating new knowledge. Furthermore, applying the orthogonality constraint to the residuals rather than to the entire experts is consistent with the goal of preserving inherited knowledge.

  • [S4] Relatively Comprehensive Experimental Evaluation

In addition to comparisons with existing continual learning and prompt-based methods, the paper includes ablation studies, limited-data experiments, and cross-backbone evaluations. The appendix also provides details on parameter settings, generator assignments, and expert construction, demonstrating a reasonable level of transparency regarding implementation details. The code is provided.

Critical weaknesses
  • [W1] It Is Unclear Whether RAMI's Improvement Is Attributable to Mechanism-Wise Task Organization (Sections 3.3, 5.3, Appendix A)

The central claim of RAMI is that organizing fake samples by generation mechanism, rather than by dataset, is beneficial. However, Table 5 in Appendix A shows that generation mechanisms and data sources are strongly correlated in the training data. Therefore, the effects of generation mechanisms and datasets are not sufficiently disentangled.

Furthermore, Protocol 4 and RAMI differ in the number of fake samples per task. According to Table 4 in Appendix A, RAMI uses 2,400 fake samples per task, whereas Protocol 4 uses different numbers of samples depending on the dataset.

Consequently, the observed performance differences may reflect several confounding factors.

  • [W2] The Central Hypothesis Behind Asymmetric Learning Is Not Directly Validated (Sections 4.1–4.4, 5.4)

RF-Prompt adopts an asymmetric design in which real knowledge is shared and protected, while fake knowledge is progressively accumulated through task-specific experts. However, the central hypothesis underlying this design—why real knowledge should be shared while fake knowledge should be separated across tasks—is not sufficiently validated through direct experiments.

Without more comparisons, it remains difficult to determine whether the improvements arise from the asymmetric architecture itself or simply from increasing the number of experts or introducing additional regularization.

  • [W3] Single-Seed Experiments Do Not Establish the Reliability of the Reported Improvements (Section 5.1)

All major experiments are based on a single run using random seed 2026.

  • [W4] Parameter-Space Orthogonality Does Not Guarantee Functional Knowledge Separation (Section 4.3, Appendix D)

The proposed method encourages residuals from different tasks to be orthogonal by minimizing an objective of the form:

$ \mathcal L_{\mathrm{orth}} \propto \left| Q_{\Delta,t}Q_{\Delta,<t}^{\top} \right|_F^2. $

However, orthogonality in the prompt parameter space does not necessarily imply independence in the model's outputs or internal representations. For example, consider two orthogonal prompt perturbations. Using a first-order approximation, their effects on the model's output features can be written as:

$ \delta f_i \approx J_f\delta p_i. $

Their functional overlap is then approximately:

$ \langle\delta f_1,\delta f_2\rangle \approx \delta p_1^\top J_f^\top J_f\delta p_2. $

In general, even if

$ \delta p_1^\top\delta p_2=0, $

it does not follow that

$ \delta p_1^\top J_f^\top J_f\delta p_2=0. $

Therefore, it remains unclear whether orthogonal residuals genuinely separate knowledge associated with different fake generation mechanisms.

Questions for authors
  1. What is the primary factor driving the improvements observed under RAMI?

  2. Is the asymmetric architecture fundamentally necessary? Have the authors compared the proposed architecture with a reversed configuration using a shared fake prompt and expanding real experts, or with symmetric architectures?

  3. Are the results reproducible across multiple random seeds?

  4. Do orthogonal residuals encode functionally distinct knowledge? Have the authors examined whether the observed parameter-space orthogonality also leads to separation in representation space or output space?

  5. What are the primary sources of forgetting? Given that historical experts are frozen, to what extent is the remaining forgetting attributable to updates to the shared real prompt, changes in the AASIST backend, or modifications to expert fusion?

  6. Does the method maintain its performance as the number of tasks increases?

Please correct me if I misunderstand something :)

4 comments
AuthorsOpenReview ↗

Response to Reviewer fYKZ - Part 1: Q1-Q3

Thank you for the detailed analysis. Following your suggestions, we added experiments on protocol attribution, asymmetric capacity allocation, and seed stability.

W1 / Q1. Disentangling optimization and source effects

We added separate controls for optimization-budget differences and mechanism/source correlation. Table 2 fixes RF-Prompt to compare organizations; the new experiments examine these alternative explanations directly.

Added optimization control. Primary epochs traverse every training row, but task sizes affect update counts and class proportions. We fix 50 epochs/task, 200 updates/epoch, and 16 real plus 16 fake utterances/batch, retaining the pools, real arrival, model, and development-based selection (seed 2026).

fake organizationCommon Avg EER (%)Pool EER (%)
Dataset-wise (P4)11.44511.830
Random four groups11.65811.972
Mechanism-wise (RAMI)10.42210.738

RAMI retains its advantage with matched updates and batch class balance (Table 4). Random groups also test mechanism organization against arbitrary equal-sized fake groups.

Added source control. Both streams use identical AT-ADD Track 2 M2/M4 utterances and real samples; only fake grouping changes. Both use 50 epochs/task, 75 updates/epoch, and 16+16 utterances/batch (seed 2026).

fake organizationCommon Avg EER (%)Pool EER (%)
Random two groups5.7505.450
M2 then M45.3205.125

Dataset identity is held fixed, so selecting different source datasets cannot explain this improvement (Table 16). Together, the controls address optimization and cross-dataset effects separately and support mechanism organization (Appendix I).

W2 / Q2. Why share real knowledge and expand fake knowledge?

The allocation follows a practical detector-update scenario. We have real speech from known domains and finite deepfake samples from observed methods, while new methods continually emerge. We protect and adapt reusable real knowledge in a shared prompt and expand fake experts for new patterns.

Related motivation already cited in Section 2:

  • ASDG (Xie et al., 2024): aggregate real representations across domains and separate fake representations.
  • OC-Softmax (Zhang et al., 2021): encourage a compact real representation.
  • LSR/LSA (Huang et al., 2025): use one real prototype and multiple fake prototypes.

Following your suggestion, we added four structure comparisons (Table 12, seed 2026). Shared branches use five tokens/layer; expanding branches add five tokens/task/layer.

real branchfake branchPool EER (%)
SharedShared13.835
ExpandingShared12.289
ExpandingExpanding12.865
SharedExpanding (RF-Prompt)10.370

The reverse has the same nominal prompt capacity; dual expansion stores more experts. Expert count alone does not explain RF-Prompt's gain. Table 3 separately evaluates protection losses. Both branches receive binary supervision; their names denote intended roles, not class-exclusive representations.

W3 / Q3. Reproducibility across seeds

We added two seeds for every method. Table 1 reports three matched seeds and mean +/- sample SD:

MethodPool EER (%)
RAWM11.692 +/- 0.361
SinglePrompt12.151 +/- 0.125
RF-Prompt10.402 +/- 0.076

Avg EER and AF also include sample SD. Partitions and task order remain fixed; supplementary controls use seed 2026.

AuthorsOpenReview ↗

Response to Reviewer fYKZ - Part 2: Q4-Q6

W4 / Q4. Parameter orthogonality and functional effects

We regularize the new adaptation directions while retaining inherited knowledge. For Pt=Bt+ΔtP_t=B_t+\Delta_t, the frozen base BtB_t transfers reusable structure. Full-expert orthogonality also penalizes inherited overlap and may encourage cancellation of transferable components; residual regularization encourages complementary increments (Figure 7, Appendix D.2).

Training evidence (seed 2026). The original loss ablation retains orthogonal initialization. Following your suggestion, we also added a full-expert/residual comparison with parent selection, inherited base, initialization, routing, and training matched.

fake regularizationPool EER (%)AF (%)
No residual loss (Table 3)12.8057.320
Full-expert orthogonality (Table 13)11.8056.560
Residual orthogonality10.3704.560

Added functional intervention. We remove residual content while retaining inherited bases, each input's original routing, the shared prompt, and backend (Table 20).

InterventionM3 EER (%)M4 EER (%)
Original checkpoint13.128.52
Remove Δ4\Delta_47.7214.20

Δ4\Delta_4 contributes to newly acquired M4 detection. Effects are task-dependent, so parameter orthogonality is not treated as a functional-independence guarantee. Training controls establish regularization benefits; interventions characterize the learned updates. Earlier residual information may persist in later frozen bases.

Q5. Forgetting despite frozen historical experts

The shared real prompt, backend, and effective mixture can still change. Following your suggestion, we added checkpoint-restoration experiments on the same 10,000 M3 utterances, without retraining (Table 11).

M3 acquisition EER after Task 3 is 4.84%. Each intervention below starts from Task 4:

Task-4 interventionM3 EER (%)
None13.12
Restore Task-3 real prompt8.88
Restore Task-3 backend12.82
Restore both shared modules8.42
Historical-only fusion; renormalized weights8.52
Restore both + historical-only fusion4.84

These interventions locate sensitivity in shared computation and fusion despite frozen experts. Their interacting effects are not additive contributions (Appendix C.3).

Q6. More tasks and recurring mechanism families

Following your suggestion, we added generator-disjoint 10/20-task streams. They retain the global pools and use matched 30,000-update budgets. The 20-task stream splits each 10-task batch into two; new generators can arrive within previously encountered families.

Seed 2026; common four-group metrics (Table 5):

TasksMethodCommon Avg EER (%)Pool EER (%)
10RAWM12.36512.505
10RF-Prompt10.42010.595
20RAWM12.24512.560
20RF-Prompt10.73510.910

RF-Prompt retains lower final detection errors at both lengths. RAWM has lower native AF. Appendix J reports assignments and native/common metrics; Appendix J.2 reports parameter/storage costs.

Thank you for the detailed rebuttal and substantial revisions. I appreciate the authors' efforts to address my concerns through additional controlled experiments. The new results significantly strengthen the empirical evidence for the proposed method. Therefore, I have changed my score from 1 to 2.

W1 / Q1: Attribution of RAMI's improvements

The fixed-step comparisons and same-source AT-ADD experiments are valuable additions. In particular, the controlled comparison among dataset-wise, random, and mechanism-wise task organizations provides stronger evidence that RAMI's improvement is not merely attributable to differences in optimization budgets or dataset sources.

However, the same-source experiment is conducted with a single seed and one random grouping, and the protocol comparisons are performed using RF-Prompt. Therefore, while the results support the effectiveness of mechanism-wise organization for RF-Prompt under the tested settings, its general effectiveness across different models, task orders, and random groupings remains uncertain.

W2 / Q2: Necessity of the asymmetric architecture

The newly added comparisons with reversed and symmetric architectures directly address my original concern. The superior performance of the proposed configuration provides convincing empirical support for the asymmetric capacity allocation.

Although these experiments do not fully establish the intended semantic distinction between real and fake knowledge, I consider the architectural design choice substantially better justified.

W3 / Q3: Reproducibility

Although the standard deviation is not always informative for experiments with only three seeds, the additional experiments with three random seeds and the reported standard deviations address my concern regarding single-seed evaluation of the main results.

W4 / Q4: Parameter orthogonality and functional knowledge separation

The matched comparison between full-expert and residual orthogonality provides useful evidence that regularizing residuals is more effective than regularizing complete experts.

I also appreciate the newly added functional interventions. However, the results reveal an important distinction between parameter-space orthogonality and functional knowledge separation.

In particular, removing the newest residual Δ4 improves M3 EER from 13.12% to 7.72%, while worsening M4 EER from 8.52% to 14.20%. Furthermore, Appendix K.1 reports that the M4 expert receives approximately 95.7% of the routing weight on M3 examples.

These observations suggest substantial functional interaction between experts associated with different tasks, despite residual orthogonality regularization.

I agree with the authors' clarification that parameter orthogonality should not be interpreted as guaranteeing functional independence. Nevertheless, I encourage the authors to maintain this distinction throughout the paper. The experiments support the empirical benefit of residual regularization, but do not establish functional separation of task-specific knowledge.

Q5–Q6: Forgetting and longer task streams

The checkpoint-restoration experiments provide a useful explanation of how forgetting can arise through changes in shared modules and expert fusion, despite frozen historical experts.

The 10- and 20-task experiments also strengthen the scalability evaluation. I appreciate the authors' acknowledgment that RF-Prompt achieves better final detection accuracy than RAWM while exhibiting greater average forgetting in these longer streams.

Overall assessment

The additional experiments address most of my original methodological concerns and considerably improve the paper.

My main remaining reservation concerns the interpretation of the results, particularly the extent to which mechanism-wise task organization generalizes beyond the tested configurations and whether residual orthogonality can be interpreted as separating task-specific knowledge at the functional level.

These limitations do not negate the demonstrated empirical improvements, but I believe the corresponding claims should be stated with appropriate qualifications.

Thank you again for the thorough and constructive response.

AuthorsOpenReview ↗

Follow-up response and manuscript clarifications

Thank you for your careful reassessment and for recognizing the additional evidence. We have now explicitly limited the protocol conclusions to RF-Prompt under the evaluated settings and clarified that residual regularization encourages complementary parameter updates without establishing functional independence (Sections 4.3, 5.5, and Conclusion). These wording changes are highlighted in blue. We appreciate your constructive feedback.

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

The paper proposes RF-Prompt for continual audio deepfake detection. It introduces a mechanism-incremental learning protocol (RAMI) and uses a shared real prompt and task-specific fake prompts to reduce forgetting. Historical experts are reused through residual learning and adaptive fusion. Experiments show improvements over several continual learning baselines.

Critical strengths
  1. The paper addresses an important problem of continual learning for audio deepfake detection.
  2. The proposed RAMI protocol provides a useful perspective.
  3. The experiments include several continual learning baselines, component ablations.
Critical weaknesses
  1. All main experiments use only one random seed. The reliability of the reported improvements is insufficient.
  2. The RAMI protocol assumes that real speech from all source domains is available from the beginning.
  3. The experiments contain only four incremental tasks. It is unclear whether the method can scale to longer task sequences as the number of fake experts grows.
Questions for authors
  1. Can the authors provide results with multiple random seeds?
  2. Can the authors evaluate longer task sequences, such as 10 or 20 tasks, and report the costs?
  3. How much of the improvement comes from the RAMI task organization? This should be discussed.
1 comment
AuthorsOpenReview ↗

Response to Reviewer fLRg

Thank you for highlighting the importance of continual ADD and the RAMI perspective. Following your suggestions, we added seed and longer-stream experiments and clarified the protocol comparison.

W1 / Q1. Multiple random seeds

We added two seeds for each compared method. Table 1 now reports mean +/- sample SD over seeds 2026/2027/2028:

MethodAvg EER (%)Pool EER (%)
RAWM11.363 +/- 0.28411.692 +/- 0.361
SinglePrompt11.698 +/- 0.10312.151 +/- 0.125
RF-Prompt10.118 +/- 0.05810.402 +/- 0.076

RF-Prompt has the lowest mean Avg and Pool EER among evaluated continual methods. Its mean AF is 4.612 +/- 0.132%, versus 5.684 +/- 0.517% for RAWM. Supplementary diagnostics use seed 2026.

W2. Availability of real-speech domains

RAMI models updating a detector with an existing multi-domain real-speech pool while new deepfake methods arrive. Recurring real-source coverage is the intended deployment setting.

Real utterances are disjoint across tasks, so historical training utterances are not replayed. Protocols 1-3 separately examine changing real distributions; Section 3 and Appendix A clarify these complementary regimes.

W3 / Q2. Ten/twenty tasks and costs

Following your suggestion, we added 10- and 20-task experiments. They retain the original sample pools and introduce disjoint generator batches within recurring mechanism families. Both methods receive 30,000 updates at each length (seed 2026).

TasksRAWM Pool EER (%)RF-Prompt Pool EER (%)RF-Prompt bank (MB)
1012.50510.5959.830
2012.56010.91019.661

RF-Prompt retains lower final detection errors at both lengths. Table 5 also reports common-group EER and native AF; Appendix J gives assignments and native averages.

The storage column counts only the float32 complete-expert-plus-base bank, not the entire model. Injected length remains ten tokens/layer, while bank storage and routing grow with tasks (Table 19, Appendix J.2).

Q3. Method gains versus protocol gains

Table 2 is the primary evidence for the protocol benefit: it fixes RF-Prompt and the global pools. P4 and RAMI share the same four-domain real-arrival strategy; fake tasks are dataset-wise and mechanism-wise, respectively.

Primary full-pass schedule, seed 2026:

ProtocolCommon Avg EER (%)Pool EER (%)
P411.03511.330
RAMI10.11010.370

RAMI gives the lowest aggregate errors among the five organizations. Complementarily, Table 1 fixes RAMI to assess method gains.

To further address your question, we added matched-budget and same-source controls as additional validation of this comparison (Tables 4 and 16).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and revision summary

We thank the reviewers for recognizing the practical relevance and coherent design of our work.

Our paper addresses detector updates using real speech from known domains and finite observed deepfake data, while new methods continually emerge. We study task organization and adaptation jointly: RAMI organizes updates by generation mechanism; RF-Prompt combines protected shared real knowledge with expanding fake experts.

Following the reviewers' suggestions, we added evidence addressing the shared concerns below.

1. Reliability (all three reviewers)

Added: two seeds per compared method. Table 1 now reports means and sample SD over seeds 2026/2027/2028.

Finding: RF-Prompt has the lowest mean Avg/Pool EERs among evaluated continual methods: 10.118/10.402%. Its pooled EER is 10.402 +/- 0.076%, versus 11.692 +/- 0.361% for RAWM, strengthening the evidence beyond the original single run.

2. Protocol attribution (fLRg W2/Q3; fYKZ W1/Q1; DcSR W1)

Primary evidence: Table 2 fixes RF-Prompt and the pools. With the same real arrival, changing P4's dataset-wise fake tasks to RAMI's mechanism-wise tasks reduces pooled EER from 11.330% to 10.370% (primary schedule, seed 2026).

Added controls: with matched updates and batch balance, P4/random/RAMI pooled EERs are 11.830/11.972/10.738% (Table 4). Within identical AT-ADD subsets, mechanism/random grouping gives 5.125/5.450% (Table 16).

These controls separately address optimization-budget and cross-dataset effects. Section 3 and Appendix A clarify RAMI's known-real-domain scenario and the changing-real-source alternatives.

3. Asymmetric capacity allocation (fYKZ W2/Q2; DcSR W3)

Added: shared/shared, reversed, dual-expanding, and RF-Prompt comparisons (Table 12). RF-Prompt achieves 10.370% pooled EER, versus 12.289% for reversal with the same nominal capacity and 12.865% for dual expansion with more experts.

Expert count alone does not explain the gain. Added freezing and class-conditioned sensitivity experiments support protected shared adaptation and the branch's stronger influence on real-example scores (Appendix K).

4. Residual learning and retention (fYKZ W4/Q4-Q5)

Added matched comparison: inheritance, initialization, routing, and training are held fixed. Residual rather than full-expert regularization reduces pooled EER from 11.805% to 10.370% (Figure 7, Table 13), supporting complementary updates that retain useful inherited structure.

Added diagnostics: fixed-routing interventions reveal task-dependent residual contributions (Table 20). Checkpoint restoration explains M3 forgetting through shared-module and fusion changes and recovers its 4.84% acquisition EER (Table 11). We distinguish parameter geometry from functional effects.

5. Longer task streams (all three reviewers)

Added: generator-disjoint 10/20-task streams with recurring mechanism families and matched update budgets. RF-Prompt/RAWM pooled EERs are 10.595/12.505% at 10 tasks and 10.910/12.560% at 20 (Table 5).

The detection advantage extends to longer streams. Appendix J reports task construction and parameter/storage costs.

New experiments, updated results, and substantive revisions are highlighted in blue in the revised PDF. Detailed responses below give the settings, results, and manuscript locations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.