TL;DR
This paper introduces the first proactive, instance-level auditing method for visual ML models, allowing individual data owners to verify if their specific data was used in training. This novel approach provides a quantifiable guarantee on false detection rates, addressing a critical gap in accountability and transparency for ML systems.
What Problem Does This Paper Address?
The increasing reliance on large datasets for training ML models has led to a surge in legal disputes over unauthorized data usage. While existing methods offer dataset-level auditing, they are insufficient for data owners with limited samples or those needing to audit individual data points. Furthermore, current instance-level auditing methods are passive and lack formal guarantees on false detection rates, making their results less reliable. This paper tackles this challenge by proposing a proactive, instance-level auditing method that offers a tunable and quantifiable false-detection rate, crucial for legal and ethical compliance.
Methods
The core of the proposed method lies in a two-component system:
- Data Marking Algorithm: Before publishing data, the owner generates n distinct marked versions of each data instance. These marks are carefully crafted to preserve the original data's utility while maximizing the distinctiveness between marked versions. This is achieved by solving an optimization problem that balances utility preservation (minimizing visual difference) and distinction (maximizing feature-space distance).
- Data-Use Detection Algorithm: This algorithm leverages any black-box membership inference method. It measures the "memorization" score of the published marked data instance relative to the hidden marked versions. A novel sequential hypothesis test is employed, which allows for early stopping and provides a quantifiable, tunable false-detection rate (FDR). This approach generalizes previous work and is designed to be effective even with a single data instance.
Datasets
The authors evaluated their method on several standard visual ML datasets:
- Image Classifiers: CIFAR-100, TinyImageNet, and ImageNet.
- Visual Encoders: CIFAR-100 and TinyImageNet.
- Vision-Language Models (CLIP & BLIP): Flickr30k.
Key Insights and Ideas
- Proactive Auditing with Guarantees: The paper introduces a proactive approach to instance-level auditing, which is a significant departure from passive methods. Crucially, it provides a quantifiable and tunable false-detection rate, a feature largely absent in prior instance-level auditing techniques.
- Generalization of Existing Work: The proposed method is a strict generalization of prior dataset-level auditing techniques, extending their capabilities to the more granular instance level.
- Effectiveness Across Diverse Models: The method demonstrates broad applicability, successfully auditing image classifiers, visual encoders, and complex vision-language models like CLIP and BLIP.
- Machine Unlearning Verification: The quantifiable FDR guarantee proves invaluable for verifying machine unlearning. The authors show that current approximate unlearning methods often fail to completely remove data influence, even when sacrificing model utility.
Key Results
- High True Detection Rates (TDR): The method achieves high TDRs across various visual ML models and datasets, often significantly exceeding the FDR bounds. For instance, in auditing image classifiers, TDRs were substantial even with low FDR bounds (e.g., 28.21% TDR at 5% FDR for CIFAR-100).
- Robustness to Countermeasures: The auditing method remains effective even when ML practitioners employ countermeasures like image perturbation, denoising, or regularization, though TDR may decrease. However, these countermeasures often significantly degrade model utility.
- Failure of Approximate Unlearning: Experiments revealed that two state-of-the-art approximate unlearning methods failed to remove the influence of data instances from image classifiers and CLIP models, as evidenced by TDRs remaining significantly higher than the FDR bound, even with a 10% drop in model utility.
Limitations and Future Work
- Computational Overhead: While the method is effective, the data marking and detection processes involve computational overhead, particularly with a large number of marked instances (n). The authors explore the impact of n and suggest that a balance between auditability and cost is necessary.
- Domain Extension: The current work focuses on visual ML models. The authors suggest extending the instance-level data-use auditing method to other domains, such as text data, as a future research direction.
- Theoretical Relationship with Differential Privacy: The paper notes that auditing differential privacy guarantees using their method is a potential avenue for future work.
Summary updated