Robustness of DINOv3 after fine-tuning

Establish whether the robustness advantage of the frozen DINOv3 feature extractor persists after task-specific fine-tuning.

Background

The benchmark evaluates DINOv3 only as a frozen feature extractor with a linear classification head, whereas the other model families are trained in different settings. Consequently, the reported degraded-set advantage does not determine whether DINOv3 would retain its robustness after adapting the backbone to the face-forgery detection task. Comparing frozen and fine-tuned configurations is needed to separate robustness attributable to the pretrained representation from robustness arising under task-specific adaptation.

References

Whether Infra-Bench-CLS-fine-tuned FMs retain their in-distribution performance under these stronger shifts is a natural extension.

— Infra-Bench CLS: A Global, Open-Source Benchmark for Critical Infrastructure Classification with Earth Observation Foundation Models  (2609.09482 - Guthrie et al., 8 Sep 2026) in Section Future work, second extension

Third, DINOv3 is evaluated only as a frozen feature extractor with a linear head, so the experiments do not establish whether its robustness advantage persists after fine-tuning.

— Benchmarking Spatial, Spectral, and Self-Supervised Cues for Face Forgery Detection under Realistic Degradation  (2609.01511 - Cunha et al., 1 Sep 2026) in Section 5, “Limitations”