TL;DR
This paper introduces a novel "frequency-based" attack that effectively bypasses existing model ownership protection methods by subtly altering a model's high-frequency components. To counter this, they propose a robust frequency-domain fingerprinting defense that leverages low-frequency characteristics for reliable ownership verification.
What Problem Does This Paper Address?
Developing high-performance deep neural networks (DNNs) is incredibly resource-intensive, involving significant investment in data, expertise, and computing power. Consequently, protecting the intellectual property (IP) of these trained models from theft and unauthorized use is a critical concern. Existing methods for model ownership resolution (MOR) primarily fall into two categories: model watermarking and model fingerprinting. While these techniques aim to prove ownership, they often struggle against sophisticated removal attacks that go beyond simple weight modifications. Specifically, current defenses are vulnerable to attacks that subtly alter model weights without significantly impacting performance, making it difficult to detect unauthorized use. This paper addresses the gap in robustly protecting DNN IP against novel removal attack vectors and proposes a more resilient defense mechanism.
Methods
The core of this paper lies in its novel approach to both attack and defense, centered around the Discrete Cosine Transform (DCT).
Frequency-based Removal Attack: The authors observe that modifying the high-frequency coefficients of a model's weights, after transforming them into the frequency domain using DCT, has a minimal impact on the model's performance. However, this modification significantly alters the model's decision boundary and weights, rendering existing fingerprinting and watermarking methods ineffective. The attack involves:
- DCT Transformation: Applying DCT to the convolutional layer weights of a DNN to convert them into the frequency domain.
- High-Frequency Coefficient Modification: Identifying and altering the high-frequency coefficients (those with higher index sums in the DCT output) to a specific value.
- Inverse DCT (IDCT): Transforming the modified frequency domain coefficients back to the spatial domain, resulting in a new set of model weights.
Frequency-based Fingerprinting Defense: To counter the proposed attack and existing removal methods, the authors propose a defense mechanism that leverages the low-frequency components of the model's weights.
- Frequency Domain Transformation: Similar to the attack, DCT is applied to the weights of a chosen layer (typically the front convolutional layer).
- Low/Mid-Frequency Component Extraction: The low and mid-frequency components (those with lower index sums) are extracted as the model's fingerprint.
- Fingerprint Generation: This extracted feature vector is then bound to the owner's identity using a random projection, creating a unique and registered fingerprint.
- Ownership Verification: When ownership is disputed, the fingerprint of a suspect model is extracted and compared (using cosine similarity) to the registered fingerprint. A high similarity indicates potential infringement.
Datasets
The paper evaluates its proposed methods on several standard image classification datasets:
- MNIST
- Fashion MNIST (FMNIST)
- CIFAR-10
- CIFAR-100
Key Insights and Ideas
- Frequency Domain as an Attack Surface: The paper reveals that the frequency domain of model weights offers a novel and effective attack surface for model ownership removal.
- High-Frequency Coefficients are Vulnerable: Modifying high-frequency DCT coefficients is a potent way to disrupt ownership protection mechanisms without degrading model performance.
- Low-Frequency Coefficients are Robust: Conversely, low-frequency DCT coefficients are more stable and retain crucial model characteristics, making them ideal for robust fingerprinting.
- DCT's Role in Concentration: Higher-dimensional DCT (like 3D DCT) better concentrates primary features in low-frequency components, minimizing performance impact when high-frequency coefficients are altered.
- Generalizability of the Approach: The frequency-based approach is shown to be effective against various existing MOR methods, including those based on decision boundaries and weights.
Key Results
- The proposed frequency-based removal attack successfully defeated state-of-the-art white-box watermarking and fingerprinting schemes while preserving model performance.
- The frequency-based fingerprinting defense demonstrated high robustness against the proposed attack and existing general removal attacks (weight pruning, filter pruning, fine-tuning).
- Fingerprint similarity (FS) for the proposed defense remained high (over 97%) across various models and datasets, even under pruning attacks that partially removed convolutional kernels.
- The proposed method is computationally efficient, with generation and verification times significantly lower than some decision-boundary-based methods.
Limitations and Future Work
- Limited Applicability to Structured Weights: The current method relies on the spatial or sequential structure of weights, making it most applicable to architectures like CNNs, ResNets, and LSTMs. Its applicability to models with less structured weights is limited.
- Future Work: The authors suggest extending their approach to support models with less spatial or sequential structure as a direction for future research.
Summary updated