Manipulating Visually-aware Federated Recommender Systems and Its Countermeasures

Wei Yuan, Shilong Yuan, Chaoqun Yang, Quoc Viet Hung Nguyen, Hongzhi Yin

Introduction

Federated Recommender Systems (FedRecs) have emerged as a powerful solution for providing personalized recommendations while maintaining user privacy. By training models on decentralized data without ever accessing raw user interactions, FedRecs mitigate many privacy concerns. However, these systems face a significant challenge: data sparsity. To combat this, modern FedRecs are increasingly "visually aware," incorporating product images into their recommendation logic to anchor item features and improve accuracy.

While the addition of visual information makes these systems more robust against traditional gradient-based hacks, it inadvertently opens a new "backdoor" for adversarial manipulation. This paper investigates the threat of image poisoning in visually-aware FedRecs and proposes a novel defense mechanism to ensure system integrity.

The Threat: PSMU++ Hybrid Attack

The authors identify a critical shift in the attack landscape. In traditional FedRecs, malicious actors upload "poisoned" gradients to promote specific items. In visually-aware systems, the visual signal acts as a "ground truth," making it harder to manipulate the model through gradients alone. This leads to the development of two new attack vectors:

  1. PSMU(V): This method generates adversarial perturbations for images. By constructing "synthetic malicious users," an attacker calculates precise modifications to an image. These modifications are human-imperceptible but cause the FedRec model's visual extractor to misinterpret the item, boosting its rank.
  2. PSMU++: This is a sophisticated "collaboration" attack. it simultaneously launches model poisoning (fake gradients) and image poisoning (adversarial images). Because both components share the same goal, they amplify each other, making the attack significantly faster and more effective than either method in isolation.
The image poisoning attack process, where synthetic users are used to construct perturbations that transform an original image into a poisoned one.

The Defense: GDMPD

To counter these potent attacks, the researchers introduced the Guided Diffusion Model for Purification and Detection (GDMPD). This defense operates at the central server level and acts as a pre-processing plug-in, requiring no changes to the underlying recommendation architecture. It functions on two primary fronts:

1. Purification

Inspired by the success of Denoising Diffusion Probabilistic Models (DDPM), GDMPD subjects every uploaded image to a "diffusion" process. This involves adding Gaussian noise to the image and then performing a "reverse" process to remove it. This "laundry machine" effect effectively washes away delicate, high-frequency adversarial noise. Because the diffusion model is trained to recover natural images, it naturally pulls the poisoned image back toward the manifold of clean, legitimate data while preserving its semantic content.

2. Detection

Beyond purification, the system must identify malicious providers. GDMPD compares the feature vectors of an image before and after the purification process using cosine similarity. If the change is significant (exceeding a pre-defined threshold), the image is flagged as adversarial. This allows the server to detect and potentially block malicious item providers.

System overview showing the interaction between benign users, attackers (model and image poisoning), and the central server protected by the GDMPD defense.

Experimental Evaluation

The researchers evaluated their methods on two widely used recommendation benchmarks: MovieLens-1M (ML), a dense dataset, and Amazon Cell Phone (AZ), a sparse dataset. They also used ImageNet to calibrate detection thresholds.

Key Results

  • Attack Potency: In visually-aware FedRecs, traditional gradient attacks were largely ineffective. However, the PSMU++ attack achieved an Exposure Rate (ER@5) of 1.0—meaning the target item appeared in the top 5 for all users—in as few as 5 to 10 training epochs.
  • Defense Efficacy: The GDMPD defense was remarkably successful. It reduced the attack's ER@5 from 1.0 down to 0.0 across both datasets and multiple model architectures, such as Fed-VNCF and Fed-LightVGCN.
  • Performance Bonus: Interestingly, using purified images slightly improved general recommendation accuracy (NDCG@20). The authors suggest the diffusion process helps by "shrinking the variance" of images from various sources, effectively cleaning up natural noise in the dataset.

Conclusion and Future Work

Visual information in Federated Recommender Systems is a double-edged sword. While it provides a robust anchor for features, it shifts the attack surface from the user to the item provider. The GDMPD defense demonstrates that diffusion models are a powerful tool for maintaining security in these systems without sacrificing performance.

Future work will focus on refining detection thresholds ($\rho$) and exploring "black-box" attacks where the adversary does not know the specific visual extractor used by the system. Additionally, optimizing the computational overhead of the 1000-step diffusion process remains a priority for scaling to massive production environments.

Summary updated

Abstract

Federated recommender systems (FedRecs) have been widely explored recently due to their capability to safeguard user data privacy. These systems enable a central server to collaboratively learn recommendation models by sharing public parameters with clients, providing privacy-preserving solutions. However, this collaborative approach also creates a vulnerability that allows adversaries to manipulate FedRecs. Existing works on FedRec security already reveal that items can easily be promoted by malicious users via model poisoning attacks, but all of them mainly focus on FedRecs with only collaborative information (i.e., user–item interactions). We contend that these attacks are effective primarily due to the data sparsity of collaborative signals. In light of this, we propose a method to address data sparsity and model poisoning threats by incorporating product visual information. Intriguingly, our empirical findings demonstrate that the inclusion of visual information renders all existing model poisoning attacks ineffective.Nevertheless, the integration of visual information also introduces a new avenue for adversaries to manipulate federated recommender systems, as this information typically originates from external sources. To assess such threats, we propose a novel form of poisoning attack tailored for visually aware FedRecs, namely image poisoning attacks, where adversaries can gradually modify the uploaded image with human-unaware perturbations to manipulate item ranks during the FedRecs’ training process. Moreover, we provide empirical evidence showcasing a heightened threat when image poisoning attacks are combined with model poisoning attacks, resulting in easier manipulation of the federated recommendation systems. To ensure the safe utilization of visual information, we employ a diffusion model in visually aware FedRecs to purify each uploaded image and detect the adversarial images. Extensive experiments conducted with two FedRecs on two datasets demonstrate the effectiveness and generalization of our proposed attacks and defenses.
BibTeX
@article{yuan2024manipulating,
  author = {Wei Yuan and Shilong Yuan and Chaoqun Yang and Quoc Viet Hung Nguyen and Hongzhi Yin},
  title = {{Manipulating Visually-aware Federated Recommender Systems and Its Countermeasures}},
  journal = {ACM Trans. Information Systems},
  year = {2024},
  url = {https://openreview.net/forum?id=biSRgBnaXd},
}
Figure