In standard Differential Privacy (DP), a single privacy budget ($\epsilon$) is typically applied to an entire dataset. However, real-world users have diverse needs. A user visiting a public park might be comfortable with low privacy, while the same user at home might demand high anonymity. This realization has led to the development of Heterogeneous Differential Privacy (HDP) and Personalized Local Differential Privacy (PLDP), frameworks that allow users to specify their own privacy levels.
The irony of these personalized systems is that the choice of a privacy level can itself be a sensitive signal. If an adversary knows that a user has requested maximum protection, they can infer that the underlying data is particularly sensitive.
The Privacy Parameter as a Side Channel
When a system allows for individualized privacy, the specific parameter chosen by a user—such as a proximity threshold in a social network or a privacy budget in a machine learning model—becomes a piece of metadata. Alaggan et al. and Gambs argue in Heterogeneous Differential Privacy that these configuration choices are not just binary settings but sensitive metadata that can reveal a user's profile. If the analyst or service provider can see these parameters, the settings act as a proxy for the very information the user intended to hide.
This leakage manifests in several ways:
- Behavioral Fingerprinting: In vehicular networks, Xu et al. et al. show in PPLS: a privacy-preserving location-sharing scheme in mobile online social networks that customized sharing preferences can act as unique fingerprints, allowing users to be tracked across supposedly anonymized datasets.
- Location Importance: Hiraishi et al. et al. demonstrate in Mechanisms to Address Different Privacy Requirements for Users and Locations that if an adversary knows a high-protection setting is active, they can infer the current location is a sensitive site, such as a home or workplace, even if the exact coordinates are obscured.
- Service Profiling: Beyond identity, the specific priorities or types of services requested by a user can be used to profile their behavior, as noted by Saxena et al. et al. in BAS-VAS: A Novel Secure Protocol for Value Added Service Delivery to Mobile Devices.
The Correlation Problem
The most significant threat occurs when a user's data and their privacy demands are correlated. For example, high-income individuals or members of vulnerable groups might systematically demand more anonymity.
Syomantak Chaudhuri explores this in Private Estimation When Data and Privacy Demands Are Correlated, noting that if an adversary knows the privacy demands of individuals, they can infer characteristics about the underlying data. This is further formalized by Chaudhuri et al. et al. in Managing Correlations in Data and Privacy Demand, where they introduce Add-remove Heterogeneous Differential Privacy (AHDP). This framework treats the privacy preference itself as a sensitive attribute that must be protected alongside the data to prevent the choice of protection from signaling the presence of sensitive characteristics.
Similarly, in the context of machine learning, Boenisch et al. et al. point out in Have it your way: Individualized Privacy Assignment for DP-SGD that a user's chosen privacy budget in a training set can act as a side channel revealing their membership in a specific group.
Mechanisms for Parameter Confidentiality
To address this, researchers are moving toward models where privacy parameters are either hidden or mathematically protected.
- Local Execution and Obfuscation: Hiraishi et al. et al. suggest that maintaining privacy mechanisms and their parameters exclusively on the user's device is a primary defense. Alaggan et al. and Gambs propose "privacy-parameter obfuscation" to ensure that the specific weights or levels chosen by a user are protected from disclosure to the analyst.
- Protecting Sensitivity Metadata: Acharya et al. et al. identify in Context Aware Local Differential Privacy that a major open question is how to protect the privacy of the sensitivity metadata itself. They argue that the mere fact that a user considers a specific data point sensitive is private information.
- Interaction Pattern Hiding: In collaborative filtering, Kim et al. et al. emphasize in Efficient Privacy-Preserving Matrix Factorization via Fully Homomorphic Encryption that protecting data values is insufficient if the set of items a user interacts with remains visible. They use Fully Homomorphic Encryption to obfuscate these interaction patterns.
- Universal Mechanisms: Sun et al. et al. note in Sum Estimation under Personalized Local Differential Privacy that while current personalized models often assume privacy levels are public, future protocols must allow for accurate aggregate estimation (like sums and means) without requiring the analyst to know each user's specific $\epsilon$.
By treating the privacy configuration as being just as sensitive as the raw data, these new frameworks ensure that the act of seeking protection does not inadvertently become a tool for deanonymization.
Key Foundational Work
Heterogeneous Differential Privacy (2016) — contributed 3 concept elements to this direction.
BAS-VAS: A Novel Secure Protocol for Value Added Service Delivery to Mobile Devices (2020) — contributed 2 concept elements to this direction.
Context Aware Local Differential Privacy (2020) — contributed 1 concept elements to this direction.