Semantic Leakage from Image Embeddings

Yiyi Chen, Anqi Wu, Desmond Eliott, Qiongxiu Li, Johannes Bjerva

Abstract: Image embeddings are generally assumed to pose limited privacy risk. We challenge this assumption by formalizing semantic leakage as the ability to recover semantic structures from compressed image embeddings. Surprisingly, we show that semantic leakage does not require exact reconstruction of the original image. Preserving local semantic neighborhoods under embedding alignment is sufficient to expose the intrinsic vulnerability of image embeddings. Crucially, this preserved neighborhood structure allows semantic information to propagate through a sequence of lossy mappings. Based on this conjecture, we propose Semantic Leakage from Image Embeddings (SLImE), a lightweight inference framework that reveals semantic information from standalone compressed image embeddings, incorporating a locally trained semantic retriever with off-the-shelf models, without training task-specific decoders. We thoroughly validate each step of the framework empirically, from aligned embeddings to retrieved tags, symbolic representations, and grammatical and coherent descriptions. We evaluate SLImE across a range of open and closed embedding models, including GEMINI, COHERE, NOMIC, and CLIP, and demonstrate consistent recovery of semantic information across diverse inference tasks. Our results reveal a fundamental vulnerability in image embeddings, whereby the preservation of semantic neighborhoods under alignment enables semantic leakage, highlighting challenges for privacy preservation.

Introduction

In the era of agentic AI and multimodal models, image embeddings have become a standard currency. These fixed-length vectors map high-dimensional images into compact spaces for efficient storage and retrieval in vector databases. Because this process is inherently "lossy"—discarding most pixel-level detail—it has been widely assumed that embeddings pose a low privacy risk.

This paper challenges that assumption. We argue that while the exact pixels of a leaked image may remain hidden, the semantic structure is highly vulnerable. The very feature that makes embeddings useful—their ability to preserve semantic similarity for retrieval—is precisely what enables privacy attacks. We introduce SLImE (Semantic Leakage from Image Embeddings), a framework demonstrating that semantic information can propagate through a sequence of lossy mappings to reconstruct detailed descriptions of private data.

The SLImE Framework

SLImE is a lightweight, multi-stage inference framework designed to exploit the intrinsic properties of embedding spaces without requiring expensive, task-specific decoders.

The SLImE framework consists of a training phase for a local retriever and an inference phase utilizing off-the-shelf models to reconstruct semantic information.

Methodology

  1. Embedding Space Alignment: An adversary aligns a "victim" embedding space with an "attacker" space using a simple linear mapping (learned via least-squares). This allows the attack to proceed even without access to the victim's specific model internals.
  2. Local Semantic Retriever: A lightweight retriever, based on a DCN v2 architecture, maps aligned embeddings to "semantic tags"—symbolic representations such as <subject, verb, object> triples.
  3. LLM/VLM Amplification: These retrieved tags are fed into off-the-shelf Large Language Models (LLMs) like DeepSeek or GPT-5 to generate coherent captions. Additionally, Vision-Language Models (VLMs) can perform "adaptive attacks," identifying objects from low-fidelity image reconstructions generated by diffusion models.

Key Insights: The "Feature" is the "Bug"

The primary objective of embedding models is to ensure that semantically similar items are clustered together. SLImE proves that this neighborhood structure is sufficient to leak private information. Even if a specific tag isn't perfectly retrieved, a "neighboring" tag is often semantically close enough for an LLM to infer the correct context. This "semantic propagation" is remarkably resilient, surviving multiple lossy transformations (Alignment $\rightarrow$ Retrieval $\rightarrow$ Discretization $\rightarrow$ LLM Generation) to produce accurate descriptions.

Experimental Results

We evaluated SLImE using the COCO and nocaps datasets across a variety of open-weights (CLIP, Nomic) and closed-API (Gemini, Cohere) models.

Alignment Efficiency and Recovery Fidelity

A critical finding is the efficiency of the attack. Significant semantic leakage occurs with as few as 1,000 to 10,000 alignment samples, a trivial requirement for a motivated adversary. As shown in the evaluation metrics, the framework can recover more than 50% of semantic information (measured by ROUGE-L) compared to human-annotated ground truth.

Plots showing the relationship between the number of alignment samples and the resulting Cosine Similarity and Rouge-L scores across different models.

Robustness and Cross-Domain Generalization

The attack proved successful across diverse models, with relation types (how objects interact) recovered with up to a 70% F1 score using Gemini-Flash. While performance dipped slightly on the nocaps dataset, the semantic structure remained partially transferable, indicating that the attack generalizes well even when the victim's images differ from the attacker's training domain.

Conclusion

Our results reveal a fundamental vulnerability in image embeddings. The preservation of semantic neighborhoods, while essential for utility, provides a direct pathway for semantic leakage. This highlights a significant challenge for privacy-preserving AI: we must develop techniques that maintain retrieval utility while intentionally disrupting the local structures that SLImE exploits. Future work will explore adaptive attacks capable of inferring even more sensitive, fine-grained details from the embedding space.

Summary updated

Abstract

Image embeddings are generally assumed to pose limited privacy risk. We challenge this assumption by formalizing semantic leakage as the ability to recover semantic structures from compressed image embeddings. Surprisingly, we show that semantic leakage does not require exact reconstruction of the original image. Preserving local semantic neighborhoods under embedding alignment is sufficient to expose the intrinsic vulnerability of image embeddings. Crucially, this preserved neighborhood structure allows semantic information to propagate through a sequence of lossy mappings. Based on this conjecture, we propose Semantic Leakage from Image Embeddings (SLImE), a lightweight inference framework that reveals semantic information from standalone compressed image embeddings, incorporating a locally trained semantic retriever with off-the-shelf models, without training task-specific decoders. We thoroughly validate each step of the framework empirically, from aligned embeddings to retrieved tags, symbolic representations, and grammatical and coherent descriptions. We evaluate SLImE across a range of open and closed embedding models, including GEMINI, COHERE, NOMIC, and CLIP, and demonstrate consistent recovery of semantic information across diverse inference tasks. Our results reveal a fundamental vulnerability in image embeddings, whereby the preservation of semantic neighborhoods under alignment enables semantic leakage, highlighting challenges for privacy preservation.1
BibTeX
@inproceedings{chen2026semantic,
  author = {Yiyi Chen and Anqi Wu and Desmond Eliott and Qiongxiu Li and Johannes Bjerva},
  title = {{Semantic Leakage from Image Embeddings}},
  booktitle = {VBN},
  year = {2026},
  doi = {10.48550/arxiv.2601.22929},
}
Figure