Towards Scalable and Robust Model Versioning

Wenxin Ding, Arjun Nitin Bhagoji, Ben Y. Zhao, Haitao Zheng

Introduction

Deep learning models represent significant intellectual property, often requiring massive investments in data curation and computational resources. Despite their value, these models are susceptible to security breaches, including server compromises and model inversion attacks. Once an attacker gains white-box access to a model, they can generate adversarial examples that bypass the system with high precision.

Traditionally, the response to a model leak is to retrain and deploy a new version. However, this strategy is undermined by attack transferability, where adversarial examples crafted for one model version remain effective against subsequent versions trained on the same dataset. This vulnerability is exacerbated by compound transferability attacks, where an adversary leverages an ensemble of all previously leaked versions to identify persistent "blind spots" in the latest replacement. This paper introduces a framework for scalable and robust model versioning designed to break this cycle of vulnerability without requiring new training data.

The Optimized Hidden Training Framework

The core of the proposed solution is Optimized Hidden Training (OHT), a technique that forces each model version to develop a unique decision boundary. The framework consists of four primary components:

  1. Parameterized Hidden Data: The training set is augmented with synthetic "hidden data." These are images perturbed such that their representations in the model's feature space cluster around a specific, task-irrelevant "hidden feature point" ($h$).
  2. Decision Boundary Distortion: By labeling this hidden data with a target class, the model is forced to incorporate these artificial features into its decision logic. This creates a unique distortion in the decision boundary that is specific to that version's hidden feature point.
  3. Greedy Optimization: Rather than selecting hidden points at random, the authors employ a greedy search algorithm. When a model is compromised, a new hidden feature point is selected from a candidate pool to specifically minimize the success rate of attacks generated by an ensemble of all previously leaked versions.
  4. GAN-Based Inversion: To ensure the hidden data does not degrade model performance, a GAN-based inversion process generates images that appear natural while mapping precisely to the desired coordinates in the feature space.

Experimental Evaluation

The authors evaluated the OHT framework across three diverse image classification tasks:

  • CIFAR10: A standard benchmark for general object recognition.
  • SkinCancer (HAM10000): A medical imaging dataset (7 classes) used to test the framework in high-stakes domains.
  • YouTubeFace: A large-scale face recognition dataset (1,283 classes) to test scalability.

Key Results

  • Utility Preservation: The models maintained standard classification accuracy across all tasks. Notably, on the SkinCancer dataset, hidden training improved accuracy from 88.7% to 90.5%, suggesting the hidden data may act as a beneficial form of data augmentation.
  • Robustness Against Compound Attacks: On CIFAR10, after eight versions were leaked, the success rate of a compound ensemble attack against the ninth version was reduced to approximately 22% (when combined with run-time filtering). In contrast, existing random-selection methods faced attack success rates exceeding 90%.
  • Operational Efficiency: Unlike robust ensemble methods that require simultaneous training of all versions, OHT allows for sequential training. Training a single version with hidden data is computationally comparable to standard training.

Discussion and Insights

The effectiveness of OHT stems from the use of task-irrelevant features as a shield. By tricking the model into relying on specific non-robust features controlled by the owner, the attacker—who lacks knowledge of the current hidden features—is forced to rely on outdated features from leaked versions.

Geometrically, the authors demonstrate through linear SVM proofs that zero transferability is theoretically possible if decision boundaries possess opposite slopes and intersect in specific feature space regions. Empirically, they found that concentrating hidden data around a single feature point per class is more effective than using multiple points, as a single point creates a sharper, more unique distortion that is harder for an ensemble attack to circumvent.

Limitations and Future Work

While OHT shows significant promise, several areas for future research remain. The current theoretical proofs are grounded in linear models; extending these proofs to Deep Neural Networks (DNNs) remains an empirical challenge. Additionally, while training is efficient, the greedy search requires generating a pool of candidate models, which introduces pre-deployment latency. Future work will also need to address adaptive attackers who might attempt to reverse-engineer the hidden distribution strategy or filter their own attacks to only include those that succeed across all known versions. Finally, the framework's performance on non-image modalities, such as text or audio, and on massive datasets like ImageNet, remains to be verified.

Summary updated

BibTeX
@inproceedings{ding2024towards,
  author = {Wenxin Ding and Arjun Nitin Bhagoji and Ben Y. Zhao and Haitao Zheng},
  title = {{Towards Scalable and Robust Model Versioning}},
  booktitle = {SaTML},
  year = {2024},
  doi = {10.1109/satml59370.2024.00036},
}
Figure