Abstract
Deep Neural Network (DNN) watermarking using backdoor triggers is a common method for intellectual property protection. However, these watermarks are often vulnerable to "Capsulation Attacks," where an adversary uses a filter to detect and block verification triggers. This paper introduces CAScore, a metric to quantify trigger detectability, and proposes a novel reverse-backdoor scheme. By utilizing normal data samples as triggers and chaining them via cryptographic hashes, the proposed method achieves near-perfect indistinguishability (CAScore of 0.99) and robust resistance against filtering attacks without compromising model performance.
Introduction
Training state-of-the-art Deep Neural Networks (DNNs) requires significant computational resources and high-quality datasets, making them valuable intellectual property. To protect these assets, model owners often embed "watermarks" using backdoor techniques. In a black-box verification setting, an owner proves ownership by demonstrating that the model produces specific, pre-defined outputs when presented with "trigger" inputs—typically images containing specific stamps or noise patterns.
However, a significant vulnerability exists: most current triggers are statistical outliers. If a pirate steals a model and suspects it is watermarked, they can implement a "Capsulation Attack."
The Capsulation Attack and CAScore
In a Capsulation Attack, the pirate places a filter (a "capsule") in front of the stolen model. This filter is designed to distinguish between normal user queries and potential verification triggers. When a trigger is detected, the filter returns a fake or neutral result, effectively "muting" the owner's evidence and preventing successful verification.
To quantify this vulnerability, the authors propose CAScore (Capsulation Attack Score). This metric is based on the Area Under the Curve (AUC) of a binary classifier trained to distinguish triggers from normal data. A CAScore of 1.0 indicates that triggers are perfectly indistinguishable from the natural data distribution, while a score near 0.0 indicates they are trivial to filter.
The Reverse-Backdoor Scheme
To counter the Capsulation Attack, the authors propose a "reverse-backdoor" scheme. Instead of using artificial, "weird" triggers, this method selects normal samples from the training set to serve as triggers.
Methodology
The scheme relies on three key components:
- Encoding: A pseudorandom inverse generator and one-way hash functions are used to chain normal samples together.
- Verification: During ownership proof, the owner submits these normal-looking triggers in a randomized sequence. Because the triggers belong to the natural data distribution, a pirate's filter cannot block them without also degrading the model's performance on legitimate user queries.
- Merkle Trees: To ensure the integrity of the trigger set, Merkle hashes provide a compact and unforgeable proof of identity.
Experimental Evaluation
The authors evaluated the proposed scheme using standard benchmarks, including MNIST and CIFAR-10, with ResNet architectures.
Key Results
- Indistinguishability: Traditional watermarking methods (e.g., noise, steganography) performed poorly on the CAScore test, with scores between 0.00 and 0.08. In contrast, the Reverse-Backdoor scheme achieved a near-perfect 0.99.
- Robustness: Under a Capsulation Attack using a Bayes filter, traditional schemes failed to verify ownership. The Reverse-Backdoor triggers successfully bypassed the filter.
- Performance: The watermarking process did not significantly impact the primary task accuracy of the models, maintaining performance levels comparable to unwatermarked baselines.
Conclusion and Future Work
The Reverse-Backdoor scheme effectively addresses the Capsulation Attack by hiding triggers "in plain sight" within the normal data distribution. While this provides strong protection in black-box scenarios, the authors note that "white-box" attacks (where the pirate has access to model weights) and "overwriting" attacks (where a pirate attempts to claim prior ownership) remain areas for future research. The use of distributed ledgers for timestamping is suggested as a potential mitigation for the latter.
Summary updated