On the (In)security of Peer-to-Peer Decentralized Machine Learning

Dario Pasquini, Mathilde Raynal, Carmela Troncoso

TL;DR

Decentralized Machine Learning (DML) promises enhanced privacy and efficiency over Federated Learning (FL) by removing the central server. However, this paper reveals that DML actually increases privacy risks. Novel attacks demonstrate that DML adversaries can be as powerful as FL servers, and achieving comparable privacy to FL requires fully connected networks, negating DML's core advantages.

What Problem Does This Paper Address?

Federated Learning (FL) trains models collaboratively across distributed devices, but relies on a central server that can be a bottleneck and a privacy risk. Decentralized Machine Learning (DML) aims to overcome these limitations by enabling direct peer-to-peer communication between clients, eliminating the central server. Proponents claim DML offers better bandwidth efficiency, user control, and crucially, enhanced privacy. This paper directly challenges the privacy claims of DML, investigating whether decentralization truly leads to better user privacy compared to FL.

Methods

The authors introduce a suite of novel privacy attacks specifically designed for DML, covering both passive and active adversarial models.

  • Passive Attacks:
    • Membership Inference Attacks (MIA): The adversary tries to determine if a specific data sample was part of a user's training set. The paper quantifies vulnerability using MIA accuracy.
  • Gradient Inversion: The adversary attempts to reconstruct training data samples from model updates (gradients). They adapt existing FL gradient inversion attacks to the DML setting.
  • Functional Marginalization: A novel attack that leverages DML's peer-to-peer nature to isolate a victim's specific contribution to their model update, enhancing MIA effectiveness.
  • Active Attacks:
    • State-Override Attack: The adversary manipulates a victim's local model aggregation process to gain complete control over their model parameters, enabling powerful inference attacks.
  • Echo Attack: The adversary forces a victim's model to overfit by repeatedly feeding back its own (or marginalized) updates, amplifying local generalization and thus privacy leakage.

The paper also analyzes the impact of communication topologies (torus, random regular, social network) and model generalization on privacy.

Datasets

The authors used the following standard datasets for their experiments:

  • CIFAR-10
  • CIFAR-100
  • STL10
  • IMDb reviews (for NLP task)

Key Insights and Ideas

  • Decentralization Doesn't Equal Privacy: Contrary to popular belief, DML does not inherently offer better privacy than FL. In fact, it often degrades it.
  • Increased Attack Surface: Removing the central server in DML creates more potential adversaries (any user can be an attacker) and increases the attack surface.
  • Local Generalization is a Privacy Leak: DML's peer-to-peer nature leads to "local generalization," where models are heavily influenced by immediate neighbors. This uneven propagation of information makes model updates more revealing about individual users' data compared to FL's global aggregation.
  • System Knowledge is Power: DML adversaries gain "system knowledge" by observing multiple neighbors' updates, allowing them to combine information and isolate individual contributions, which is not possible in FL.
  • Active Adversaries in DML are Potent: Active DML adversaries can achieve the same level of control as a malicious FL server, even overriding a victim's local model aggregation.
  • Trade-off Between Privacy and Efficiency: Defenses that improve DML privacy (like fully connected networks) eliminate its efficiency advantages over FL.

Key Results

  • Passive Attacks: DML adversaries can achieve higher MIA accuracy than FL adversaries, especially when the DML topology is sparse. The functional marginalization attack significantly boosts MIA effectiveness.
  • Active Attacks: The state-override attack allows DML adversaries to gain full control over victim models, enabling powerful gradient inversion attacks comparable to those by FL servers. The echo attack effectively increases generalization error and privacy leakage.
  • Topology Matters: Sparse topologies in DML exacerbate privacy risks due to increased local generalization. Dense topologies reduce these risks but also reduce efficiency gains.
  • No Free Lunch: Achieving privacy comparable to FL in DML requires fully connected networks, negating the benefits of decentralization.

Limitations and Future Work

  • Defenses: The paper discusses defenses like Secure Aggregation (SA) and Differential Privacy (DP). While SA can mitigate some attacks, it introduces significant overhead and can be evaded in DML. DP in DML (local-DP) offers a worse privacy-utility trade-off than central-DP in FL. Robust aggregation methods can be bypassed or even amplify privacy leakage.
  • Future Work: The authors suggest that more research is needed to find effective and reliable topology-aware SA in DML. They also highlight the need for new design principles to enable truly privacy-preserving DML. The paper also notes that while their attacks are demonstrated on specific architectures and datasets, they are generally applicable to DML.

Summary updated

Abstract

In this work, we carry out the first, in-depth, privacy analysis of Decentralized Learning—a collaborative machine learning framework aimed at addressing the main limitations of federated learning. We introduce a suite of novel attacks for both passive and active decentralized adversaries. We demonstrate that, contrary to what is claimed by decentralized learning proposers, decentralized learning does not offer any security advantage over federated learning. Rather, it increases the attack surface enabling any user in the system to perform privacy attacks such as gradient inversion, and even gain full control over honest users’ local model. We also show that, given the state of the art in protections, privacy-preserving configurations of decentralized learning require fully connected networks, losing any practical advantage over the federated setup and therefore completely defeating the objective of the decentralized approach.
BibTeX
@inproceedings{pasquini2023security,
  author = {Dario Pasquini and Mathilde Raynal and Carmela Troncoso},
  title = {{On the (In)security of Peer-to-Peer Decentralized Machine Learning}},
  booktitle = {SP},
  year = {2023},
  doi = {10.1109/sp46215.2023.10179291},
}
Figure