Research map / Continual learning / Continual fine-tuning of LLMs

DevFD : Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces

NeurIPS 2025 · Continual learning › Continual fine-tuning of LLMs · approach: Synthetic data and merging for CLIP · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Frames face forgery detection as continual learning and proposes DevFD, a developmental mixture-of-LoRA-experts with one shared Real-LoRA and orthogonal Fake-LoRAs trained with an integrated orthogonal (subspace + gradient) loss and a label-guided balancing loss.

Problem

Face forgery detectors handle known manipulation types but forget earlier forgery types and struggle to adapt to new ones when trained sequentially under limited data and compute.

Approach

A developmental MoE of LoRA experts: a single shared Real-LoRA models the stable real-face distribution, while a growing sequence of Fake-LoRAs is kept orthogonal to previous subspaces via an orthogonal loss augmented with an orthogonal-gradient term, plus a label-guided localized balancing loss that routes real vs fake samples.

Key results

• Under the dataset incremental protocol, DevFD reaches 93.48 average accuracy... with average forgetting 1.35 on DFDC-P, and an overall average of 89.82 vs DMP's 89.99 (DMP is best on the final average).
• Under the manipulation types incremental protocol (AUC), DevFD achieves 94.90 average AUC with 2.44 forgetting, tied with SUR-LID at 94.90 on average AUC but with lower forgetting than SUR-LID's 1.26 being lower; DevFD reports the best final-task performance.
• Trainable parameters are about 1.18% of total at the first task and 2.89% after four tasks; the authors claim roughly 1/10 the trainable parameters of compared methods.

Measured effects reported in the paper

Surprising findings

• Real-LoRA is not shown to cause forgetting; replacing it with an orthogonal sequence gives no significant gain.
• Adding only the orthogonal loss without gradient term already yields large forgetting reduction, while the paper claims gradient term further reduces forgetting in early training.

Limitations

• Model size grows linearly with the number of tasks as Fake-LoRAs are added.
• Relies on a strong pre-trained backbone that already encodes general knowledge.
• No error bars or repeated runs due to limited resources.

Experimental setup

Datasets: FaceForensics++ DFDC-P DFD Celeb-DF v2 DF40

Benchmarks: FaceForensics++ DFDC-P DFD Celeb-DF v2 DF40

Metrics: accuracy AUC average accuracy average forgetting

Baselines: LwF ER SI CoReD DFIL DMP iCaRL DER HDP SUR-LID

Builds on: LoRA O-LoRA InfLoRA DFIL MoECL

Closest papers by meaning

• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.88)
• Stacking Brick by Brick: Aligned Feature Isolation for Incremental Face Forgery Detection (CVPR 2025 · similarity 0.88)
• Agent4FaceForgery: Multi-Agent LLM Framework for Realistic Face Forgery Detection (CVPR 2026 · similarity 0.87)
• Forensics Adapter: Adapting CLIP for Generalizable Face Forgery Detection (CVPR 2025 · similarity 0.87)
• Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.87)
• FakeDiffer: Distributional Disparity Learning on Differentiated Reconstruction for Face Forgery Detection (AAAI 2025 · similarity 0.86)