SIF: Semantically In-Distribution Fingerprints
for Large Vision-Language Models

University of Central Florida

Abstract

TL;DR — Existing LVLM fingerprints are easy to detect and erase because they are semantically abnormal. SIF keeps the ownership signal in-distribution, achieving state-of-the-art stealthiness and robustness with no parameter changes.

The public accessibility of Large Vision–Language Models (LVLMs) raises concerns about unauthorized reuse. Existing ownership-verification methods rely on semantically abnormal queries or out-of-distribution responses, which adversaries can easily detect and remove. We first expose this via the Semantic Divergence Attack (SDA), which filters fingerprint checks by measuring semantic divergence from a reference model. To address it, we propose SIF (Semantically In-Distribution Fingerprints), a non-intrusive framework. Semantic-Aligned Fingerprint Distillation (SAFD) distills text-generation watermark signals into the visual modality for semantically coherent yet fingerprinted responses, and Robust-Fingerprint Optimization (RFO) simulates worst-case representation perturbations for resilience to fine-tuning and quantization. Experiments on LLaVA-1.5 and Qwen2.5-VL show that SIF achieves superior stealthiness and robustness.

The Vulnerability: Semantic Divergence Attack

Prior fingerprints are semantically abnormal by design, so a stealer can run a small reference LVLM beside the stolen model: abnormal prompts (high perplexity) or outputs that diverge from the reference are flagged and replaced with the reference response — scrubbing the fingerprint.

Comparison between existing LVLM fingerprinting methods and SIF.
Existing methods — IF, ProFLingo, and PLA — rely on unnatural triggers or semantically irrelevant responses (e.g. answering “CVPR Conference” for a skier). SIF instead produces in-distribution, semantically coherent fingerprint responses.
Semantic Divergence Attack: a reference model flags suspicious queries and responses.
The Semantic Divergence Attack measures semantic divergence from a benign reference model, flags suspicious queries and responses, and substitutes the reference output — removing semantically abnormal fingerprints.

SIF: Hiding the Fingerprint in Plain Sight

SIF pipeline: fingerprint construction with SAFD and RFO, then black-box copyright verification.
(a) A tiny perturbation distills a text watermark into a trigger image (SAFD), hardened against representation drift (RFO). (b) Standard decoding on the suspect API emits the signal — verified fully black-box.

Semantic-Aligned Fingerprint Distillation

Distills a decoding-based text watermark into the input image. A small perturbation (16/255) is optimized so the model generates green-list watermark tokens while keeping the response close to a natural teacher output — so the reply is natural yet carries a detectable signal.

Robust-Fingerprint Optimization

Stealers quantize or fine-tune models, which shifts internal representations and breaks fragile fingerprints. RFO simulates worst-case representation perturbations during optimization to make the trigger more robust, so the fingerprint survives such model modifications.

Verification example: the trigger image elicits watermarked responses from suspect models (detected) but not from an unrelated model.
Suspect models are detected via a strong watermark signal, while an unrelated model is not.

Robustness to Model Modifications

In practice, a stolen model is often modified before deployment. Across quantization (4-/8-bit) and full fine-tuning on six downstream datasets, SIF consistently retains the highest fingerprint matching rate on both LLaVA-1.5-7B and Qwen2.5-VL-7B, while prior methods are largely erased.

Robustness table: Fingerprint Matching Rate under quantization and full fine-tuning, SIF vs. baselines.
Fingerprint Matching Rate (FMR, higher is better) under quantization and full fine-tuning. SIF (Ours) achieves the best robustness across every setting. “N/A”: IF requires fingerprint embedding before model release.

BibTeX

@InProceedings{Zhao_2026_CVPR,
    author    = {Zhao, Yifei and Lou, Qian and Zheng, Mengxin},
    title     = {SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {17399-17408}
}