Bayesian Perspective on Memorization and Reconstruction

2025-05-29

AI Summary

This paper addresses the formal characterization and defense against data reconstruction attacks, with a focus on clarifying the fundamental nature of fingerprinting code (FPC) attacks. Adopting a Bayesian perspective, the authors provide the first formal definition of reconstruction attacks and rigorously distinguish them from membership inference attacks—demonstrating that FPC attacks are inherently membership inference, not reconstruction. Building on this framework, they introduce a provably secure definition for reconstruction-specific defenses and show that classical “reconstruction impossibility” results no longer hold when defenses target reconstruction alone (i.e., without addressing membership inference). The main contributions are threefold: (1) the first Bayesian formalization of the reconstruction problem; (2) a corrected theoretical classification of FPC attacks; and (3) a novel, feasible defense paradigm targeting reconstruction attacks exclusively—enabling finer-grained privacy objectives in mechanism design.

Technology Category

• Machine Learning: Privacy
• Computer Vision: Adversarial Attacks & Robustness
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

We introduce a new Bayesian perspective on the concept of data reconstruction, and leverage this viewpoint to propose a new security definition that, in certain settings, provably prevents reconstruction attacks. We use our paradigm to shed new light on one of the most notorious attacks in the privacy and memorization literature - fingerprinting code attacks (FPC). We argue that these attacks are really a form of membership inference attacks, rather than reconstruction attacks. Furthermore, we show that if the goal is solely to prevent reconstruction (but not membership inference), then in some cases the impossibility results derived from FPC no longer apply.

Problem

• Introduces Bayesian perspective on data reconstruction attacks
• Reinterprets fingerprinting code attacks as membership inference
• Shows reconstruction prevention possible despite FPC impossibility results

Innovation

• Bayesian perspective on data reconstruction
• New security definition prevents reconstruction attacks
• Reinterprets fingerprinting code as membership inference