EDGE-Shield: Efficient Denoising-staGE Shield for Violative Content Filtering via Scalable Reference-Based Matching

Authors: Takara Taniguchi, Ryohei Shimizu, Duc Minh Vo, Kota Izumi, Shiqi Yang, Teppei Suzuki

Published: 2026-04-04 08:32:24+00:00

AI Summary

EDGE-Shield proposes a scalable, reference-based content filter to mitigate copyright violation and deepfake generation by Text-to-Image models. It operates during the denoising process, leveraging embedding-based matching for efficiency and an x-pred transformation to enable accurate detection at early stages. The method significantly reduces processing time for violative content filtering while maintaining accuracy across different generative models.

Abstract

The advent of Text-to-Image generative models poses significant risks of copyright violation and deepfake generation. Since the rapid proliferation of new copyrighted works and private individuals constantly emerges, reference-based training-free content filters are essential for providing up-to-date protection without the constraints of a fixed knowledge cutoff. However, existing reference-based approaches often lack scalability when handling numerous references and require waiting for finishing image generation. To solve these problems, we propose EDGE-Shield, a scalable content filter during the denoising process that maintains practical latency while effectively blocking violative content. We leverage embedding-based matching for efficient reference comparison. Additionally, we introduce an \\textit{$x$}-pred transformation that converts the model's noisy intermediate latent into the pseudo-estimated clean latent at the later stage, enhancing classification accuracy of violative content at earlier denoising stages. We conduct experiments of violative content filtering against two generative models including Z-Image-Turbo and Qwen-Image. EDGE-Shield significantly outperforms traditional reference-based methods in terms of latency; it achieves an approximate $79\\%$ reduction in processing time for Z-Image-Turbo and approximate $50\\%$ reduction for Qwen-Image, maintaining the filtering accuracy across different model architectures.

Key findings

EDGE-Shield significantly reduces filtering latency, achieving approximately a 79% reduction for Z-Image-Turbo and 50% for Qwen-Image, while maintaining high classification accuracy (ROC-AUC ~0.85). It demonstrates superior scalability, exhibiting near-constant inference latency even with an increasing number of reference images, and its x-pred transformation enables accurate violative content detection at early denoising stages.

Approach

EDGE-Shield operates by pre-computing and caching embeddings of reference images for efficient comparison. During the Text-to-Image model's denoising process, it applies an x-pred transformation to the noisy intermediate latent to estimate a pseudo-clean latent. This estimated latent is then decoded, its embedding is extracted, and a similarity score against the cached references determines if the generation should be halted due to violative content.

Datasets

HUB dataset, CPDM dataset

Model(s)

UNKNOWN

Author countries

Japan

← Previous