TL;DR
This paper introduces PGAC, a novel framework for crafting adversarial attacks against Vertical Federated Learning (VFL) systems. PGAC enables an attacker to dominate joint inference without needing access to the server model, real test data, or training data, by leveraging auxiliary data from non-training domains.
What Problem Does This Paper Address?
Vertical Federated Learning (VFL) allows multiple parties to collaboratively train a machine learning model without sharing their raw data, by leveraging complementary feature spaces. This is particularly useful in scenarios where different entities possess different attributes of the same entities (e.g., a bank having credit history and a retailer having purchase history for the same customer). However, VFL systems are vulnerable to adversarial attacks. Existing attacks often require extensive knowledge of the server model, access to real test samples, or labeled data from the training domain, which are often impractical in real-world VFL deployments due to privacy concerns and resource limitations. This paper addresses the need for a practical and effective adversarial attack framework that operates under more relaxed assumptions, specifically targeting the joint inference phase of VFL.
Methods
PGAC is a two-module framework:
- Proxy Generation: This module aims to create a high-fidelity proxy model that mimics the server model's behavior. It achieves this by:
- Generalized Embedding Learning: Using domain adversarial networks to learn domain-invariant features from auxiliary (source) and target (attacker's) data. An attention mechanism is incorporated to emphasize category-specific information.
- Domain Knowledge Transfer: Employing a multi-task discriminator to align source and target embeddings, while down-weighting outlier source classes using class transferable probabilities. This helps in transferring knowledge from auxiliary data to the target domain effectively.
- Self-training and Gradient Similarity Loss: Further refining the proxy model by using pseudo-labels for unlabeled target data and minimizing the discrepancy between the proxy's and server's gradients.
- Adversarial Crafting: This module uses the generated proxy model to craft adversarial inputs that dominate joint inference. It involves:
- Shadow Input Construction: Creating a diverse set of "shadow inputs" by emphasizing salient regions in auxiliary images (identified using Grad-CAM) and then scaling and duplicating them. These shadow inputs simulate influential test inputs from other parties.
- Dominant Perturbation Generation: Iteratively generating perturbations on the attacker's input to minimize the loss for a desired target class while simultaneously reducing the contribution of the shadow inputs (representing benign parties) to the joint inference. This is achieved using a gradient-based optimization method (Projected Gradient Descent) with momentum.
Datasets
The paper evaluates PGAC on three cross-domain image classification datasets:
- Office-31: Contains 4,652 images across 31 categories from three domains (Amazon, DSLR, Webcam).
- ImageNet-Caltech: Built using 84 shared classes between ImageNet and Caltech-256.
- DomainNet: Utilizes four domains (Clipart, Painting, Real, Sketch) with 126 classes, focusing on the first 40 alphabetically.
The VFL configuration typically involves two parties, each owning half of the feature vector. The attacker uses auxiliary data from non-training domains, which can have different distributions and class spaces than the VFL training data.
Key Insights and Ideas
- Leveraging Auxiliary Data: The core innovation is the effective use of auxiliary data from non-training domains to build a powerful proxy model, bypassing the need for real training data or server queries.
- Proxy Fidelity is Crucial: A high-fidelity proxy model is essential for accurately guiding adversarial perturbations, making the attack transferable to the actual server model.
- Shadow Inputs for Realistic Simulation: Creating diverse shadow inputs that mimic benign parties' contributions is key to crafting adversarial inputs that can effectively "monopolize" rewards by suppressing others.
- Cross-Domain Adaptation is Solvable: The paper demonstrates that even with significant cross-domain disparities, knowledge transfer for adversarial purposes is feasible.
- Practicality Under Relaxed Assumptions: PGAC's strength lies in its practicality, requiring minimal attacker knowledge, which is a significant step forward for adversarial attacks in VFL.
Key Results
- PGAC achieves a high attack success rate (ASR), with results approaching the "Real Data" benchmark (which assumes full knowledge) and significantly outperforming existing state-of-the-art attacks. For example, on Office-31, PGAC achieved an ASR of 96.33% with VGG-16 models.
- The framework demonstrates strong performance across various embedding and server model architectures, indicating its generality.
- Increasing the volume of auxiliary data improves attack performance, but the gains stabilize after a certain point. Even with a relatively small amount of auxiliary data (100 samples per class), PGAC achieves high ASR and fidelity.
- PGAC remains effective even with an increased number of benign passive parties (up to 32).
- The paper also shows that PGAC generally survives existing defenses, though stringent settings for private training can mitigate its effectiveness at the cost of main task accuracy.
Limitations and Future Work
- Defense Evasion: While PGAC shows resilience against common defenses, the paper notes that stringent settings for private training can counter it, albeit with a significant drop in classification accuracy. This suggests that future work could explore more robust defense mechanisms.
- Auxiliary Data Requirement: Although PGAC doesn't require training data, it does rely on labeled auxiliary data from non-training domains. The availability and quality of this auxiliary data could impact attack performance.
- Computational Cost: Crafting adversarial inputs, especially with methods like PGD, can be computationally intensive. The paper doesn't extensively discuss the computational trade-offs.
- Future Directions: The authors suggest exploring more sophisticated countermeasures and investigating the impact of different types of auxiliary data and VFL configurations. Further research into adaptive defenses against PGAC would also be valuable.
Summary updated