FedOT: Ownership Verification and Leakage Tracing via Watermarks for Federated LDMs

2026-06-22

AI Summary

This work addresses the vulnerability of shared latent diffusion models (LDMs) in federated learning to unauthorized redistribution by malicious clients, a risk exacerbated by the inadequacy of existing watermarking techniques in enabling reliable source tracing and their susceptibility to VAE replacement attacks. To overcome these limitations, the authors propose FedOT, the first framework to embed traceable watermarks in federated LDMs. FedOT employs block-wise watermarking to simultaneously support model ownership verification and identification of malicious clients, while introducing a latent vector transformation (LVT) module to enhance robustness against VAE replacement attacks. Experimental results demonstrate that FedOT not only enables effective ownership validation and precise leakage source attribution but also significantly degrades the visual quality of images generated by tampered models, thereby achieving a strong balance between security and usability.

Abstract

Training Latent Diffusion Models (LDMs) within Federated Learning (FL) has attracted increasing attention due to its ability to combine the powerful generative capacity of LDMs with the privacy-preserving properties of FL. However, FL requires sharing the global model with multiple participants, which risks unauthorized model distribution or resale by malicious clients. While an intuitive approach is to adopt existing VAE-based watermarking techniques for LDMs in FL, this strategy falls short in addressing such threats due to two fundamental challenges: (1) Existing methods support ownership verification but lack the ability to trace model leakage to a specific malicious client; (2) VAE-based watermarks are vulnerable, as they can be removed simply by replacing the decoder with a clean counterpart. In this paper, we propose FedOT, the first framework for ownership verification and leakage tracing in federated LDMs. Specifically, to address the first challenge, we design a chunked watermark, where the first part is for ownership verification, and the second part is used for client identification. Furthermore, to overcome the second challenge and secure the model against VAE replacement attack, we introduce Latent Vector Transformation (LVT), which strengthens the connection between the VAE and U-Net latent spaces by modifying the original latent distribution of the VAE. Consequently, any attempt to replace the VAE for watermark removal leads to significant image quality degradation, making the LDM model unusable. Extensive experiments demonstrate that FedOT achieves superior performance in both ownership verification and traceability. Project page: https://spyzixuan.github.io/FedOT/.

Problem

• Federated Learning
• Latent Diffusion Models
• Model Watermarking
• Ownership Verification
• Leakage Tracing

Innovation

• Federated Learning
• Latent Diffusion Models
• Watermarking
• Ownership Verification
• Leakage Tracing