Parameter-matched LoRA placement reveals performance–efficiency trade-offs in cross-dataset deepfake detection
Cross-dataset generalization remains difficult in deepfake detection, and the placement of parameter-efficient adaptation within pretrained vision models remains insufficiently controlled. We study standard weight-space Low-Rank Adaptation (LoRA) in CLIP ViT-L/14 from two complementary perspectives. A primary comparison evaluated attention output (Attn-Out), attention query/value (Attn-Q/V), MLP, and MLP+Attn-Out using exactly 1,478,658 trainable parameters each and three independent seeds. A common-rank contrast showed that Attn-Out used 49.9% fewer trainable parameters than Attn-Q/V at \(r=16\) , while its source-validation frame-level AUC was lower by 0.000659. Models were trained on FaceForensics++ C23, and configurations were frozen using source validation before evaluation on six reporting-only external datasets. In the exact-budget comparison, Attn-Q/V achieved the highest observed mean Macro-6 AUC ( \(0.9432\,\pm\,0.0010\) ), whereas Attn-Out led numerically on three targets and required less measured training memory and model-compute time. All family-balanced crossed-bootstrap intervals included zero, so the AUC ordering is descriptive. Explicit merging removed placement-specific low-rank operations, and inference measurements did not support a deployment-speed ranking. Together, these comparisons separate adapter size and training-resource requirements from merged inference behavior.
Authors
- Hyunsoo Kang
- Dohyeon Kim
Institutions
- Chungbuk National University (KR)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1038/s41598-026-80070-7
- Primary Topic
- Digital Media Forensic Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00