A Secure Fingerprinting Framework for Distributed Image Classification

Guowen Xu, Xingshuo Han, Anguo Zhang, Tianwei Zhang

TL;DR

This paper introduces SECUREMARK-DL, a novel framework for securing deep learning image classification models in distributed settings. It uniquely embeds customer-specific fingerprints for intellectual property protection and traitor tracking, while simultaneously safeguarding training data privacy through a new privacy partitioning technique.

What Problem Does This Paper Address?

The widespread commercialization of deep learning (DL) models, particularly in image classification, has created a significant need for intellectual property (IP) protection. Current methods often focus on embedding watermarks into models to detect copyright infringement. However, these approaches fall short in two critical areas: they rarely address the problem of "traitor tracking" – identifying authorized customers who misuse or redistribute the models – and the process of embedding watermarks in distributed training environments can compromise the privacy of the sensitive data used for training. This paper aims to bridge this gap by providing a unified framework that tackles both IP protection with traitor tracking and data privacy in distributed DL.

Methods

SECUREMARK-DL employs two primary techniques:

  • Privacy Partitioning: The DL model is split into two parts: a local network ($\Theta_L$) distributed to data owners and a remote network ($\Theta_R$) kept on the cloud server. Data owners train their local network with their private data, and an adversarial DNN is used to obfuscate the intermediate outputs, preventing data leakage. This obfuscated data is then used to train the complete model collaboratively.
  • Unique Fingerprinting with Traitor Tracking: For each customer, a unique fingerprint ($F_i$) is generated and embedded into the $\Theta_R$ part of the model. This fingerprint consists of two components:
  • Community Relationship Code (CRC): This code leverages social network structures (e.g., based on geographical proximity or relationships) to quickly screen potential colluding groups. It uses Neighborhood Preserving Hashing (NPH) to map high-dimensional social data into low-dimensional binary vectors.
  • Customer Identity Code (CIC): This code is generated using Balanced Incomplete Block Designs (BIBD) to uniquely identify each customer and to detect collusion among a specific number of customers. This ensures that even if multiple customers collude to average their model parameters, their combined identity code can still be identified.

The fingerprint is embedded by adding a regularization term to the model's objective function during training, minimizing the difference between the fingerprint and a projected version of the model's parameters.

Datasets

The paper evaluates SECUREMARK-DL on three standard image classification datasets:

  • MNIST
  • LFW (Labeled Faces in the Wild)
  • CIFAR-10

Key Insights and Ideas

  • Dual-Purpose Fingerprints: The ingenious design of the fingerprint, combining CRC for group screening and CIC for individual identification and collusion detection, is a key innovation. This allows for efficient identification of potential infringers and precise tracking of malicious actors.
  • Privacy-Preserving Distributed Training: The privacy partitioning method, where the model is split and adversarial DNNs are used to obfuscate intermediate data, effectively protects sensitive training data without requiring data owners to share it directly.
  • Robustness to Attacks: The framework demonstrates significant resilience against common attacks like model fine-tuning, model compression, and fingerprint collusion, maintaining high accuracy and effective fingerprint detection.
  • Trade-off Management: The paper highlights the inherent trade-off between model utility (accuracy) and privacy protection, suggesting that the placement of the partition point in the network is crucial for balancing these factors. Deeper partition points offer better privacy but might slightly reduce accuracy.

Key Results

  • High Classification Accuracy: SECUREMARK-DL maintains high classification accuracy (over 95%) even with a long-bit (304-bit) fingerprint embedded.
  • Effective Privacy Partitioning: Experiments on MNIST show that privacy partitioning significantly degrades the quality of recovered images (measured by MSE and SSIM) while preserving high model accuracy.
  • Robust Fingerprint Detection: The fingerprinting mechanism shows high detection rates (close to 100%) and low false alarm rates against various attacks, including model fine-tuning and compression. For instance, the (121, 11, 1)-BIBD CIC can detect up to 10 colluders with a 100% detection rate and 0% false alarm rate.
  • Negligible Overhead: The computational overhead for fingerprint embedding and extraction is reported to be almost negligible compared to the total training cost.

Limitations and Future Work

  • Trade-off between Privacy and Utility: The paper acknowledges that placing the partition point in deeper layers enhances privacy but can slightly compromise classification accuracy. The authors suggest that a careful trade-off is necessary in practical applications.
  • Computational Cost of Re-training: Embedding a unique fingerprint for each customer requires re-training the model, which, while efficient, is still a cost.
  • Future Work: The authors suggest future research will focus on improving the security and efficiency of SECUREMARK-DL. This could involve exploring more advanced privacy-preserving techniques or optimizing the fingerprint embedding and extraction processes.

Summary updated

Abstract

The deep learning (DL) technology has been widely used for image classification in many scenarios, e.g., face recognition and suspect tracking. Such a highly commercialized application has given rise to intellectual property protection of its DL model. To combat that, the mainstream method is to embed a unique watermark into the target model during the training process. However, existing efforts focus on detecting copyright infringement for a given model, while rarely consider the problem of traitors tracking. Moreover, the watermark embedding process can incur privacy issues for the training data in a distributed manner. In this paper, we propose SECUREMARK-DL, a novel fingerprinting framework to address the above two problems in a distributed learning environment. It embeds a unique fingerprint into the target model for each customer, which can be extracted and verified from any suspicious model once a dispute arises. In addition, it adopts a new privacy partitioning technique in the training process to protect the training data privacy. Extensive experiments demonstrate the robustness of SECUREMARK-DL against various attacks, and its high classification accuracy (> 95%) even if a long-bit (304-bit) fingerprint is embedded into an input image.
BibTeX
@misc{xu2022secure,
  author = {Guowen Xu and Xingshuo Han and Anguo Zhang and Tianwei Zhang},
  title = {{A Secure Fingerprinting Framework for Distributed Image Classification}},
  year = {2022},
  eprint = {2207.04668},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2207.04668},
}
Figure