Abstract
Visual Prompt Learning (VPL) has emerged as a highly efficient paradigm for adapting large-scale pre-trained models to downstream tasks by learning small, task-specific pixel patches. However, this efficiency may come at a significant privacy cost. This paper investigates the susceptibility of visual prompts to privacy-leaking attacks, specifically Property Inference Attacks (PIA) and Membership Inference Attacks (MIA). Our findings reveal that despite their minimal parameter count, prompts act as dense representations of training data, allowing adversaries to reconstruct macro-level dataset properties and identify individual training samples with high accuracy.
Introduction
As machine learning models scale to billions of parameters, traditional full-parameter fine-tuning has become computationally and storage-prohibitive. Visual Prompt Learning (VPL) addresses this by keeping the backbone model frozen and only optimizing a small set of parameters—the "prompt"—added to the input data. This efficiency has facilitated the "Prompt-as-a-Service" (PaaS) model, where specialized prompts are traded as proprietary assets.
The prevailing assumption was that the extreme compression inherent in VPL (where prompts often comprise less than 0.1% of the model's parameters) would naturally safeguard the privacy of the underlying training data. This research challenges that notion, demonstrating that these tiny pixel perturbations are remarkably "leaky."
Methodology
The study evaluates privacy risks through two primary attack frameworks adapted for the VPL context:
- Property Inference Attacks (PIA): These attacks aim to deduce global characteristics of the training set, such as demographic distributions (e.g., gender or age ratios). The adversary employs Shadow Prompt Generation, training their own prompts on datasets with known properties to create a meta-classifier capable of auditing target prompts.
- Membership Inference Attacks (MIA): These attacks attempt to determine if a specific individual's data was used to train a prompt. The authors explore neural network-based, metric-based (e.g., prediction entropy), and gradient-based attack vectors.
Experimental Setup
The assessment was conducted across four diverse image datasets representing various sensitive attributes:
- CIFAR10: General image classification.
- CelebA: Large-scale facial attributes for inferring gender and age ratios.
- UTKFace: Facial images labeled with gender, race, and age.
- AFAD (Asian Face Age Dataset): Focused on age and gender demographics.
Key Insights and Results
The Density of Prompt Information
Despite their small size (e.g., 69k parameters compared to 86M in a ViT-B backbone), prompts are surprisingly data-rich. Adversaries achieved 100% accuracy in identifying training dataset sizes and between 81% and 99% accuracy in inferring demographic proportions on the CelebA dataset. For membership inference, metric-based attacks reached up to 93% accuracy on the AFAD dataset.
Overfitting and Vulnerability
A critical finding is the strong correlation (Pearson score of 0.89) between a prompt's level of overfitting and its vulnerability to MIA. Prompts trained for excessive epochs or on insufficient data are significantly more likely to leak individual membership information.
The Defense-Utility Trade-off
While standard defenses like adding Gaussian noise to the prompt can mitigate membership inference, they often fail to protect against property inference without severely compromising the prompt's primary task performance.
Conclusion and Future Work
This research highlights a significant "privacy tax" associated with Visual Prompt Learning. The current reliance on simple noise-based defenses is insufficient for protecting macro-level dataset properties. Future research must explore more robust privacy-preserving techniques, such as integrating Differential Privacy (DP-SGD) directly into the prompt-learning phase, and investigating black-box attack scenarios where the underlying model architecture is unknown to the adversary.
Summary updated