Research map / Multi-agent systems and games / LLM agents

Agent4FaceForgery: Multi-Agent LLM Framework for Realistic Face Forgery Detection

CVPR 2026 · Multi-agent systems and games › LLM agents · approach: Agent safety and attack benchmarks · 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.

Proposes Agent4FaceForgery, an LLM multi-agent simulation (profile, memory, action modules plus social roles and adaptive rejection sampling) that generates multimodal face-forgery data with text-image consistency labels to train detectors.

Problem

Face forgery detectors trained on static benchmark data generalize poorly to real-world social media forgeries, because training data lacks diverse creator intent, iterative creation process, and social text-image interactions.

Approach

LLM-powered agents with profile (traits derived from FF++ creators), memory, and action modules simulate forgery creation; an adaptive rejection sampling gate filters samples using agent self-assessment plus an external detector; a social simulation with roles (Watcher, Explorer, Critic, Chatter, Poster, Gemini Auditor) generates misleading or accurate text to create text-image consistency labels.

Key results

• Cross-dataset frame-level AUC of 87.10% on Celeb-DF and 86.50% on WildDeepfake, top or second-best versus listed SOTA methods.
• On DF40 six-manipulation robustness, average AUC of 93.9% versus 78.7% for the best listed baseline (ProgressiveDet).
• Sequential training with A4FF data raises LLaVA AUC from 83.2% to 92.2% and lowers EER from 24.8% to 16.8%.

Surprising findings

• Ablation shows the PNS (social simulation) module alone reaches 91.0% CDF AUC, more than the ARS-only (88.0%) or FT-only (83.2%) variants.
• Diminishing returns in agent count: 12 agents gives only marginal gains over 6 (Celeb-DF AUC 85.8% vs 85.3%).

Limitations

• Forged-sample generation relies on closed API models (GPT-4V, Gemini) and detectors trained on FF++, which may bias the synthetic data.
• Only about 25k generated pairs; the paper's text is inconsistent on details such as the number of profiles, the role list, and which tables use which backbone.
• Many method details (thresholds, lambda, q, warm-up size, prompts) are not specified.

Experimental setup

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

Benchmarks: FF++ DFD DFDC-P Wild Deepfake Celeb-DF DF40

Metrics: AUC EER ACC precision recall F1 AP

Baselines: Xception EfficientNet-B4 Face X-ray F3-Net MAT GFF LTW LRL DCL SBI UIA-ViT RECCE

Builds on: AI Town AutoGen MetaGPT Agent4Rec FF++

Closest papers by meaning

• DevFD : Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces (NeurIPS 2025 · similarity 0.87)
• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.86)
• Towards General Visual-Linguistic Face Forgery Detection (CVPR 2025 · similarity 0.86)
• Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.86)
• A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.85)
• VLForgery Face Triad: Detection, Localization and Attribution via Multimodal Large Language Models (NeurIPS 2025 · similarity 0.84)