acceptodds
Under review as a conference paper at ICLR 2027

Multimodal DeepFake Detection via Domain-Incremental Multi-Adapter Learning

Abstract

DeepFake detectors are typically trained in a static setting and struggle to generalize to emerging manipulation techniques. As generative methods evolve rapidly, the ability to incrementally extend models to new domains without full retraining becomes critical. We introduce DEFEND, a framework for domain-incremental multimodal DeepFake detection with parameter-efficient updates. Unlike prior work that focuses on dataset-level shift, we consider a fine-grained incremental setting, where new manipulation types arrive sequentially. To this end, we construct a chronological stream of 23 domains, spanning from early face-swapping methods to recent multimodal forgeries. DEFEND builds on a frozen pretrained audio-visual backbone and leverages Low-Rank Adaptation (LoRA) to enable efficient domain-specific updates while mitigating catastrophic forgetting. Importantly, we introduce a dynamic adapter expansion guided by a learned key-query routing strategy. This mechanism allows the model to reuse existing adapters for similar manipulations while expanding its capacity when encountering novel ones. We evaluate DEFEND on three benchmarks, KoDF, DeepSpeak and MAVOS-DD. On BRAVEn, DEFEND reaches 89.7% average AUC under automatic routing, 8.5 points above the strongest continual learning baseline, with only 11 adapters for 23 domains; on CLIP+CLAP, it is on par with the best continual learning baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
incremental learning, domain adaptation, deepfake detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_doYYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets the domain‑incremental multimodal DeepFake detection task. In real‑world scenarios, forgery techniques keep evolving over time. Conventionally static‑trained DeepFake detectors exhibit poor generalization against emerging forgery methods. Full retraining incurs high computational costs and requires storing all historical training data. The authors propose DEFEND, which builds upon frozen pre‑trained audio‑visual backbones and leverages LoRA low‑rank adapters for parameter‑efficient incremental updates to mitigate catastrophic forgetting. Its core innovation is a learnable key‑query routing mechanism that dynamically expands the adapter set: for each incoming new forgery domain, the algorithm automatically decides whether to reuse an existing adapter or instantiate a new one. The authors construct a chronological incremental data stream consisting of 23 forgery domains covering face‑swap, lip‑sync, and avatar‑based forgery techniques spanning 2017 to 2024, sourced from KoDF, DeepSpeak, and MAVOS‑DD. Comprehensive ablation and comparison experiments are conducted on two backbones, BRAVEn and CLIP+CLAP. With only 11 adapters for all 23 domains, DEFEND achieves an average AUC of 89.7% under automatic routing on BRAVEn, outperforming traditional continual‑learning baselines by a notable margin. On the CLIP+CLAP backbone, its performance is on par with the strongest baseline. The paper further conducts analyses on routing behaviour, domain grouping, memory overhead, cross‑language performance, and other aspects.

Critical strengths

The problem formulation aligns with real world forensic scenarios. Most existing DeepFake detectors follow a static training setup; the few existing incremental detection works mostly define domains at the dataset level. This paper treats each forgery generation method as an independent domain and constructs a chronological 23 domain data stream, which reflects the real world iterative evolution of forgery techniques. The method is reasonably designed. Built upon LoRA adapters, two routing strategies are proposed: the cosine similarity based SIM routing and the learnable key query routing DEFEND KQ. The mechanism dynamically decides whether to reuse an existing adapter or instantiate a new one, eliminating the need for a dedicated adapter per domain and substantially reducing parameter overhead. A margin loss is further designed to separate features of real and fake samples and assist the learning of the routing module. The experimental suite is comprehensive. • A rich set of baselines are compared: single adapter, one adapter per domain, Joint Training, EWC, DER++, DualPrompt, and KD+Margin Loss. • Extensive ablation studies are conducted: LoRA rank, adapter placement layers, reuse margin, routing network capacity, upper bound of stored embeddings for the router, and the presence/absence of the candidate adapter compatibility score snews_{new}. • In depth analyses are provided: automatic domain grouping behaviour, random routing control experiments, cross language performance, performance comparison between audio manipulated domains and real audio only domains, and routing accuracy variation against the number of seen domains. Thorough computational overhead analysis is provided. The paper reports detailed statistics for training/inference FLOPs, wall clock time, GPU memory consumption, and parameter counts, demonstrating practical deployment potential.

Critical weaknesses
  1. The paper motivates its work by noting that forgery techniques keep evolving over time and conventional statically trained DeepFake detectors generalize poorly against novel forgery types. However, all test experiments are derived from internal splits of four datasets: KoDF, DeepSpeak, MAVOS DD and DFDC. No fully independent external datasets are adopted for out of distribution (OOD) evaluation. Although temporally later domains are unseen forgery algorithms from the model’s perspective during incremental learning, all samples still originate from these four datasets.
  2. DEFEND assumes known domain boundaries during training, where each incremental step receives clean samples belonging to a single domain. In real world online scenarios, incoming samples are unlabeled mixed forgery data: one cannot know which forgery type a sample belongs to, and samples mixed with multiple forgery patterns may even exist.
  3. The key query router must store sample embeddings for every historical domain, resulting in linearly growing memory consumption with the number of domains. While ablations over the maximum storage budget B are presented, routing AUC drops noticeably when B=500. For real world production systems that continuously ingest new forgery techniques, memory usage will grow without bound.
  4. Conventional continual learning baselines (EWC, DER++, KD) fine tune the full head module, whereas DEFEND only updates LoRA adapters. This creates mismatched parameter update granularity. Although total parameter counts are fully reported in the paper, one critical baseline is missing.
  5. Building blocks including LoRA, adapter based Mixture of Experts, and key query matching are pre existing techniques, leading to limited theoretical novelty.
  6. Multiple formatting issues exist. For instance, the first formula lacks an equation number.
  7. Visualizations illustrating how routing accuracy and average AUC change as the number of domains increases are insufficient.
Questions for authors
  1. How would DEFEND perform on a completely unseen external dataset without access to any training samples from that new domain?
  2. Could the authors add small scale mixed domain simulation experiments, where each incremental step mixes samples from multiple forgery domains, to quantify the performance degradation of DEFEND?
  3. Given unbounded memory growth for continuously incoming new forgery techniques and performance drop at B=500, what mitigation strategies can be adopted for practical deployment?
  4. Why are regularizers such as EWC / DER++ not directly applied to LoRA adapters (instead of full head fine tuning) to remove unfair architectural confounding factors? Could this set of baselines be supplemented?
  5. The core building blocks are existing techniques. Could the authors further clarify the unique non trivial contributions of this work beyond existing components?
  6. Could the authors fix formatting errors such as the missing equation number for the first formula in the revision?
  7. Could additional visualizations be added to illustrate how routing accuracy and average AUC change as the number of domains increases?
3 comments

We thank the reviewer for their feedback and for highlighting our chronological manipulation benchmark, adaptive adapter reuse, and analysis of computational costs. We address the concerns below.

Q1: DEFEND targets incremental adaptation as labeled real/fake data arrive. To assess transfer before adaptation, we evaluated BRAVEn AV and V on FakeAVCeleb [1] without using its data for training, adapter allocation, or router fitting. We evaluate BRAVEn AV and V on FakeAVCeleb [1], completely unseen to DEFEND. We compare to AVAD [2] and RealForensics [3]; for both, we compute the AUC over the entire set of manipulations in the FakeAVCeleb dataset (not originally done in the AVAD paper). For RealForensics, we pretrain the model on DeepSpeak.

MethodAll AUC (%) ↑
AVAD [2]79.16
Real Forensics [3]87.30
BRAVEn AV, DEFEND-KQ85.63 ± 0.66
BRAVEn V, DEFEND-KQ89.71 ± 1.54
Both DEFEND models transfer to the unseen FakeAVCeleb dataset, reaching 85.6 - 89.7% overall AUC, which is better than AVAD and on par with RealForensics. This experiment measures zero-shot transfer to FakeAVCeleb. DEFEND’s primary goal is incremental adaptation, where labeled real and fake clips become available over time and DEFEND can reuse or train an adapter.

Q2: See response W1 to reviewer Wuew.

Q3: Like experience replay, DEFEND trades storage for performance. The BRAVEn router stores 4 KiB per clip embedding. Capping storage at 500 embeddings per domain reduces router memory from 217 to 39.7 MiB, compared with 62.5 MiB for the DER++ replay buffer. Routed AUC drops by 1.73 points (89.72% to 87.99%, Table 15). Smaller budgets reduce adapter reuse and routing accuracy. The cap applies per domain, so total router memory still grows with the number of domains.

Q4: Thank you for the suggestion. Below, each LoRA baseline updates a single adapter, matching one DEFEND adapter.

MethodTrainable parametersBRAVEn AV AUC (%)BRAVEn AV forgettingBRAVEn V AUC (%)BRAVEn V forgetting
Fine-tuning, full head8.67M†71.82 ± 3.9127.80 ± 2.7657.91 ± 2.2127.80 ± 2.76
Fine-tuning, LoRA1.05M70.93 ± 6.6824.32 ± 6.7787.88 ± 0.837.22 ± 1.15
KD+Margin, full head8.67M†55.27 ± 3.2229.80 ± 3.9578.21 ± 1.7214.33 ± 2.11
KD+Margin, LoRA1.05M81.14 ± 5.1713.63 ± 5.4387.98 ± 1.037.34 ± 1.18
EWC8.67M81.23 ± 0.7110.65 ± 0.7885.44 ± 0.935.18 ± 1.00
EWC, LoRA1.05M81.15 ± 0.758.95 ± 0.4486.07 ± 0.312.94 ± 0.35
DER++8.67M78.51 ± 2.518.50 ± 2.5178.58 ± 2.282.18 ± 1.06
DER++, LoRA1.05M89.36 ± 1.732.65 ± 1.3785.34 ± 0.483.51 ± 0.43
DEFEND-KQ1.05M per adapter89.68 ± 0.904.24 ± 1.0587.95 ± 1.155.79 ± 1.17
On BRAVEn AV, DEFEND-KQ still exceeds KD+Margin+LoRA by 8.54 AUC points and reduces forgetting from 13.63 to 4.24. EWC+LoRA maintains similar AUC and reduces forgetting somewhat. Notably, DER++ with LoRA improves from 78.51% to 89.36% AUC and reduces forgetting from 8.50 to 2.65. It now reaches comparable AUC to DEFEND-KQ (89.36% vs. 89.68%). On BRAVEn V, DEFEND-KQ achieves similar AUC to LoRA fine-tuning and KD+Margin (87.95% vs. 87.88–87.98%) with lower forgetting (5.79 vs. 7.22–7.34). EWC and DER++ have lower forgetting (2.94 and 3.51 vs 5.79) and lower AUC (86.07 and 85.34 vs 87.95). Overall, DEFEND-KQ still reliably delivers top performance across both backbones.

Q5: DEFEND’s main contribution is a rule for deciding when to add or reuse a LoRA adapter. For each new domain, DEFEND trains a candidate adapter and keeps it if its compatibility score exceeds that of the best existing adapter by a margin. Otherwise, it discards the candidate and updates the best existing adapter. Removing the candidate comparison leaves five adapters and lowers routed AUC (Table 23). During inference, the frozen DFDC-pretrained head produces the query embedding. The router scores the query against the learned adapter keys. A second pass applies the highest-scoring adapter to classify the clip. On our chronological benchmark of 23 audio and video manipulation domains from three datasets (2017-2024), DEFEND matches the routed AUC of one adapter per domain under the same routing procedure and storage budget (89.68% vs. 89.57%) while using roughly half as many adapters (11.3 vs. 23, Reviewer EnMs, Q2). The learned groups often follow manipulation families across datasets (Table 8). For held-out domains, training on other domains from the same group yields higher AUC than training on an equal number from other groups (89.55% vs. 85.50%, Table 9).

Q7 Figure 3 and Table 10 in Appendix E show how routing accuracy and routed AUC change as new domains are added. We welcome suggestions for other ways to visualize these results.

We fixed the formatting error and updated the PDF with the revised version covering Q3.

AuthorsOpenReview ↗

References

[1] Khalid et al. (2022), FakeAVCeleb: A Novel Audio-Video Multimodal Deepfake Dataset [2] Feng et al. (2023), Self-Supervised Video Forensics by Audio-Visual Anomaly Detection [3] Haliassos et al. (2022) Leveraging Real Talking Faces via Self-Supervision for Robust Forgery Detection

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

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

The paper presents DEFEND for multimodal deepfake detection in a continual learning setting. A learned router decides whether to update an existing LoRA adapter or retain a newly trained adapter when a new manipulation domain arrives. The authors evaluate the method on a chronological stream assembled from several datasets. DEFEND performs well with BRA VEn and matches the strongest evaluated continual learning baselines with CLIP+CLAP.

Critical strengths
  1. Updating detectors as new manipulation techniques emerge is a well-motivated problem. The extended manipulation-level stream provides a useful test of whether adaptation can preserve detection performance on earlier domains.

  2. The automatic-routing results with BRA VEn provide meaningful evidence for the method. The comparison with ground-truth routing helps quantify how much detection performance is lost when domain labels are unavailable.

Critical weaknesses
  1. The router is trained on adapter-specific embeddings, but inference requires a query before an adapter is selected. The paper does not specify which model produces this query. This omission prevents the reader from reconstructing a central part of the proposed method.

  2. The independent-adapter baseline is evaluated using ground-truth domain labels, without a result using the same learned router as DEFEND. That comparison is needed to determine whether adaptive reuse improves the practical tradeoff over keeping a separate adapter for each domain.

  3. The splitting procedure does not establish whether each original video and its manipulated derivatives remain within the same split. The source distribution of real test clips is also unclear, leaving uncertainty about the influence of dataset-specific cues.

  4. Some tables are difficult to read, and the algorithm text overlaps with the surrounding prose.

Questions for authors
  1. Which model produces the query embedding before adapter selection at inference? How does this input correspond to the embeddings used during router training?

  2. How does the independent-adapter baseline perform with the same learned router and embedding-storage budget as DEFEND?

  3. Does the split keep each original video together with its manipulated derivatives? How are the sources of real test clips matched to those of the manipulated clips?

5 comments

We thank the reviewer for recognizing the value of our manipulation-level data stream, and the value of comparing automatic and ground-truth routing on BRAVEn.

Q1: During inference, to produce the embedding used to compute the routing query, we use the frozen pretrained head. The router scores the query against the learned adapter keys and selects the highest-scoring adapter. The detection head then runs with that adapter to produce the real/fake prediction.

We compared this approach to a more expensive one, using the BRAVEn AV backbone. Here, the queries come from all the different adapters, are compared to the keys, and we select the highest-scoring adapter. Across three seeds, it achieved 89.90 ± 0.52% routed AUC with 11.7 adapters, compared to 89.68 ± 0.90% with 11.3 adapters for the original “cheaper” approach. Given the comparable AUC, we use the simpler method to compute the query in a single pass before adapter selection.

Q2: As suggested, we evaluated the one-adapter-per-domain baseline (23 adapters) with a learned router on BRAVEn AV using the full training set and three seeds. Both the one-adapter-per-domain and DEFEND use the same key–query router architecture, training procedure, hyperparameters, and embedding storage budget (217 MiB).

MethodAdaptersAdapter paramsRouted AUC (%) ↑Routed forgetting ↓Routing acc. (%) ↑
One adapter per domain + KQ2324.2M89.57 ± 0.693.77 ± 0.7160.02 ± 0.53
DEFEND-KQ11.31.05M per adapter89.68 ± 0.904.24 ± 1.0577.88 ± 8.09

DEFEND achieves comparable routed AUC with roughly half the adapter parameters (89.68% vs. 89.57%; 11.9M vs. 24.2M). One adapter per domain has slightly lower forgetting, as expected. Routing accuracy is higher with DEFEND-KQ, where several domains can share the same correct adapter. We will add this comparison to Table 5 and use it in Sec 5.3 to discuss the trade-off between adaptive reuse and one adapter per domain.

Q3:

Splits. All splits are disjoint at the clip level: we verified that none of the test clips appear in the training or validation data in the incremental stream. The DFDC pretraining and test splits are also clip-disjoint. All methods in Table 5 use the same split files.

  • DeepSpeak v1/v2: We use the official, identity-disjoint splits.
  • KoDF:, DFDC, MAVOS-DD: We use a random clip-level split for these datasets, as we have no identity breakdown in the metadata.

This is also consistent with previous works [1, 2, 3], which do not focus on identity-disjoint splits but on sample-based splits.

Sources of real test clips. Each incremental domain draws its test real samples from the same dataset(s) as its fakes, as well as from DFDC. See also our response to W4, Reviewer Wuew.

[1] Oorloff T., et al,, AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (2024)

[2] Datta et al., PIA: Deepfake Detection Using Phoneme-Temporal and Identity-Dynamic Analysis (2025)

[3] Liu W, et al., Lips Are Lying: Spotting the Temporal Inconsistency between Audio and Visual in Lip-syncing DeepFakes (2024)

We updated the PDF submission by adding the requested experiment (Q2) and fixing the formatting issue. We would appreciate information on which tables are difficult to read, so we can improve them.

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

Just another kind reminder; thanks for your review!

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

This paper proposes DEFEND, a domain-incremental framework for multimodal DeepFake detection. It maintains a set of LoRA adapters over frozen audio-visual backbones and uses a learned key-query router to decide whether to reuse an existing adapter or introduce a new one as new manipulation domains arrive. At inference time, the router selects an adapter without requiring domain labels. Experiments with CLIP+CLAP and BRAVEn show that DEFEND improves detection performance over several continual learning baselines on BRAVEn.

Critical strengths

1.The paper studies domain-incremental DeepFake detection at the manipulation level, using 23 chronologically ordered domains rather than a small number of dataset-level tasks. This provides a useful setting for examining how detectors adapt to newly emerging manipulation techniques.

2.DEFEND addresses adapter selection at inference time without assuming access to domain labels. The adapter grouping analysis in Table 9 also provides evidence that related manipulation domains can benefit from sharing adapters.

Critical weaknesses

1.DEFEND assumes that manipulation domains arrive sequentially with known boundaries and labeled training data. However, real-world DeepFake streams may contain multiple manipulation types without clear domain boundaries. It remains unclear how the proposed domain-level adapter allocation could be applied when manipulation types are mixed or unknown.

2.The performance gains vary considerably across backbones. In Table 5, DEFEND substantially outperforms continual learning baselines on BRAVEn, but achieves comparable performance on CLIP+CLAP (77.96% vs. 77.91% AUC for DER++). Whether the main benefit comes from mitigating severe forgetting in particular backbone architectures rather than consistently improving domain-incremental detection?

3.The routing mechanism requires storing embeddings from previously observed domains. Table 15 shows that reducing the memory budget from all embeddings to 500 per domain increases the number of adapters from 15.7 to 21.0 and decreases routed AUC from 89.72% to 87.99%. Does it means that adapter reuse and routing performance are sensitive to the available memory?

4.The experimental protocol may introduce dataset-specific shortcuts. As acknowledged in the Conclusion, real samples are drawn from DFDC and the incremental datasets in a fixed 3:1 ratio. Whether the detector can exploit differences in data sources rather than manipulation-specific evidence?

Questions for authors

See comments above.

4 comments

We thank the reviewer for their thoughtful feedback and for recognizing the value of our manipulation-level domain-incremental setting and domain-label-free adapter selection. Below, we address the concerns.

W1: Indeed, we assume that the data comes in a segmented way: as a new video generator becomes available, we obtain a new batch of data and incrementally update our model. We see this as a realistic scenario. Moreover, during training, adapter reuse/allocation is decided once per episode, which can contain, in some cases, the same manipulation type from different sources (datasets). Inference routing is per sample and requires no domain information.

Still, to assess DEFEND’s robustness to “mixed” streams, we merge the 23 domains into Consecutive triples (chronological; 8 batches) and Random triples (mixed across datasets/years; 8 batches), without distinguishing them. All the training details and metrics match Table 5, results on BRAVEN-AV.

StreamDomainsAdaptersRoutingAvg AUC (%) ↑Forgetting ↓Forgetting w/o DFDC ↓Routing acc. (%) ↑
Original2311.3GT-KQ92.38 ± 0.731.42 ± 0.921.49 ± 0.96100
DEFEND-KQ89.68 ± 0.904.24 ± 1.053.81 ± 1.0277.88 ± 8.09
Consecutive triples86.0 ± 1.0GT-KQ94.07 ± 0.570.86 ± 0.700.90 ± 0.73100
DEFEND-KQ91.28 ± 0.563.77 ± 0.693.11 ± 0.6272.12 ± 4.45
Random triples84.3 ± 0.6GT-KQ93.10 ± 0.811.47 ± 0.961.53 ± 1.00100
DEFEND-KQ91.32 ± 0.933.33 ± 1.092.89 ± 1.1981.25 ± 4.67

Mixed triples achieve comparable routed AUC to the original stream (91.3% vs. 89.7%), with even forgetting (3.3-3.8 vs. 4.2). Each update sees multiple domains simultaneously, so an adapter sees more diverse data. Some domains have few samples, and merging them potentially gives better adapters. Consecutive and random triples see more data per update than the original stream, suggesting that the amount of data seen during adapters’ training matters. These results hint at DEFEND’s applicability to “mixed” data streams.

W2: Gains across backbones.

Indeed, performance gains vary across backbones, and backbones themselves vary in strength. For the baselines, e.g., EWC obtains 75.87% on CLIP+CLAP and 81.23% on Braven-AV; DualPrompt has a similar behaviour. One possible reason is that BRAVEn is pretrained on lip-sync and talking-head manipulations, while CLIP and CLAP learn general features by aligning images and audio to text. Generally, we report several backbones to show that DEFEND is applicable in all cases, and the improvement from DEFEND is indeed consistent. Compared to using a single LoRA adapter, DEFEND-KQ improves the avg. AUC from 74.28 to 77.96% on CLIP+CLAP, from 66.63% to 89.68% on BRAVEn AV, and from 84.32% to 87.95% on Braven V.

W3: Indeed, reducing memory affects adapter allocation and routing as shown in Table 15. A smaller router budget (B=500 samples) reduces AUC due to the router's weaker decision about which adapter to use, but overall yields about 2-point delta compared to a router trained on all samples (89.72% vs 87.99%, respectively). Memory dependency is also a factor for any Experience Replay based methods, such as DER++. At the same time, our 87.99% is 9.5 points above DER++ (78.51%), while using 39.7 MiB versus 62.5 MiB by DER++.

W4: In our paper, we consider a scenario where we might have few real examples for the current domain, for that reason, we employ the 3:1 split, where 3 parts come from the pretraining dataset (DFDC), and 1 part comes from the current domain. To test any bias introduced by the use of DFDC, we removed DFDC real samples from incremental training and validation data. Each sample was replaced with a real clip from the domain’s own dataset. Sample counts, fakes, test sets, DFDC pretraining, and hyperparameters remained unchanged. AUC is averaged over T0–T23; forgetting uses automatic routing and the Table 5 definition; we report BRAVEn AV over three seeds.

Training realsGT AUC (%) ↑Routed AUC (%) ↑Forgetting ↓Adapters
3:1 (DFDC:current domain), paper92.38 ± 0.7389.68 ± 0.904.24 ± 1.0511.3
0:1, No DFDC93.37 ± 0.5190.93 ± 0.564.57 ± 0.5110.7

Removing DFDC real samples does not impact the final results, which are overall on par with the original results. This shows that DEFEND does not exploit the DFDC-versus-incremental-dataset cue.

We have updated the PDF submission to include the requested experiments (W1 and W4).

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

Just another kind reminder. Thanks for your review!

Thanks for the additional experiments and clarification. I'll maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the reviews' responses

We thank the reviewers for the constructive feedback. We answered all three reviews, ran the requested experiments, and updated the PDF. Below, we summarize the main points:

Shared concerns

  • Mixed streams and domain boundaries (Wuew, doYY). We merged the 23 domains into 8 batches of consecutive or random triples. Routed AUC on BRAVEn-AV was 91.3% for both, versus 89.7% for the original stream, with similar forgetting. We still assume labeled data arrives in batches.
  • Dataset shortcuts and splits (Wuew, EnMs). Removing all DFDC real samples from incremental training gave 90.93% routed AUC (vs. 89.68%). All splits are clip-disjoint, and DeepSpeak uses its official identity-disjoint splits.
  • Router memory (Wuew, doYY). With 500 embeddings per domain, routed AUC drops from 89.72% to 87.99%. This is still 9.5 points above DER++ (78.51%) with less memory (39.7 vs. 62.5 MiB). Total memory still grows with the number of domains.

Reviewer-specific

  • EnMs: At inference time, the query comes from the frozen DFDC-pretrained head, and a costlier variant that computes queries through every adapter gives the same AUC (89.90% vs. 89.68%). A one-adapter-per-domain baseline with the same router and budget reaches 89.57% with 23 adapters (24.2M parameters), versus 89.68% for DEFEND with 11.3 adapters (11.9M parameters).
  • Wuew (backbones): Gains are largest on BRAVEn and on par with the best baselines on CLIP+CLAP. DEFEND still beats a single LoRA adapter on every backbone.
  • doYY: Zero-shot AUC on FakeAVCeleb is 85.6% (AV) and 89.7% (V), versus 79.2% for AVAD and 87.3% for RealForensics (neither is designed for incremental learning). We added LoRA versions of EWC, DER++, and KD+Margin. On BRAVEn-AV, DER++ with LoRA is comparable to DEFEND (89.36% vs. 89.68%). On BRAVEn-V, DEFEND achieves performance comparable to EWC and DER++. We also clarified our contribution, the add-or-reuse rule based on a compatibility score, and fixed the formatting issues. Overall, DEFEND-KQ still reliably delivers top performance across both backbones.

We updated the PDF with a revised version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.