ML Security PapersFake-HR1: Rethinking Reasoning of Vision Language Model for Synthetic Image Detection
Changjiang Jiang 1,2, Xinkuan Sha 2, Fengchang Yu 1, Jian Liu 2, Mingqi Fang 2, Chenfeng Zhang 2,3, Jingjing Liu 2, Wei Lu 1
Published on arXiv
2602.10042
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Fake-HR1 adaptively applies Chain-of-Thought reasoning only when needed, surpassing existing LLMs on generative detection benchmarks while significantly reducing token consumption and latency.
Fake-HR1 (HFT + HGRPO)
Novel technique introduced
Recent studies have demonstrated that incorporating Chain-of-Thought (CoT) reasoning into the detection process can enhance a model's ability to detect synthetic images. However, excessively lengthy reasoning incurs substantial resource overhead, including token consumption and latency, which is particularly redundant when handling obviously generated forgeries. To address this issue, we propose Fake-HR1, a large-scale hybrid-reasoning model that, to the best of our knowledge, is the first to adaptively determine whether reasoning is necessary based on the characteristics of the generative detection task. To achieve this, we design a two-stage training framework: we first perform Hybrid Fine-Tuning (HFT) for cold-start initialization, followed by online reinforcement learning with Hybrid-Reasoning Grouped Policy Optimization (HGRPO) to implicitly learn when to select an appropriate reasoning mode. Experimental results show that Fake-HR1 adaptively performs reasoning across different types of queries, surpassing existing LLMs in both reasoning ability and generative detection performance, while significantly improving response efficiency.
Key Contributions
- Hybrid Fine-Tuning (HFT) framework that partitions data into reasoning and non-reasoning subsets to cold-start a dual-mode VLM without costly manual annotation
- Hybrid-Reasoning Grouped Policy Optimization (HGRPO), an RL-based online training stage that teaches the model to implicitly select between reasoning and direct-response modes
- Fake-HR1, the first adaptive hybrid-reasoning model for AIGC detection that balances inference efficiency with detection accuracy across varying forgery complexity
🛡️ Threat Analysis
Fake-HR1 is explicitly a synthetic image / AIGC detection system — it detects AI-generated content (deepfakes, diffusion-model outputs), which is a canonical ML09 output integrity / content provenance use case.