Selective Data Sharing Mechanisms

• Selective data sharing is a framework that enables controlled disclosure of only authorized data subsets based on policies, tasks, or relevance criteria.
• It employs techniques such as cryptographic enforcement, fragmentation, and delegated computation to ensure fine-grained access control and integrity.
• These systems balance privacy and utility by optimizing efficiency trade-offs and enabling minimal data release with immediate revocation capabilities.

Searching arXiv for the cited papers and closely related selective data sharing work to ground the article. Search query: PeerShare selective data sharing arXiv (Nagy et al., 2013) Selective data sharing denotes a family of mechanisms in which only an authorized subset of data, data-derived outputs, or model-internal state is disclosed, and only under explicit policy, task, or relevance constraints. Across the literature, the selective object may be a social-graph–scoped data item, a subset of encrypted files, a minimal data capsule, a framed JSON view, a delegated computation result, a selectively perturbed table, or an experience or KV subset exchanged between learning agents. The unifying theme is that disclosure is not “all or nothing”: systems instead bind release to social relationships, attributes, tasks, capabilities, contracts, or utility–privacy objectives, and then enforce those constraints cryptographically, systemically, or both (Nagy et al., 2013, Kuo et al., 2021, Xia et al., 2023).

1. Definitions and recurring design objectives

In the systems literature, selective data sharing is usually defined operationally rather than axiomatically. In PeerShare, the core operation is to distribute “shared keys, public keys and any other data that need to be distributed with authenticity and confidentiality guarantees to an authorized set of recipients, specified in terms of social relationships” (Nagy et al., 2013). In 3LSAA, the objective is “fine-grained, owner-defined access control” and “private keyword search” on an “honest-but-curious” server, while preserving “data self-sovereignty” (Kuo et al., 2021). In Data Station, the objective is to let data owners “share data with the escrow knowing it will not be released without their consent,” with data users delegating computation and receiving only permitted outputs (Xia et al., 2023). In SelectShare, the emphasis is on “discovery and selective disclosure of IoT data without violating their integrity” and on “fine-grained access control over the shared data” (Fotiou et al., 2022).

A recurring systems-level distinction concerns what exactly is shared. Some schemes share plaintext data items under constrained readership, as in PeerShare’s AppData objects parameterized by a sharing policy (Nagy et al., 2013). Some share encrypted files discoverable only through selective search and attribute satisfaction, as in 3LSAA’s SSE–ABE–AES stack (Kuo et al., 2021). Some share only fragments, as in the MCPS method where the public cloud receives $PUF_c$ and reconstruction additionally requires $PRF_c$ from the trusted smartphone (Qiu et al., 2019). Others share a task-specific result rather than the source data itself, as in Data Station’s delegated computation model (Xia et al., 2023). Still others share only a minimal unit, as in TD-DCSS where the data capsule is “the minimal unit of storage/share” (Lyu et al., 2024).

A second recurring distinction concerns the selection criterion. In several papers, selection is policy-driven: social lists in PeerShare, attributes and owner-issued trapdoors in 3LSAA, capabilities in SelectShare, or contracts in Programmable Dataflows (Nagy et al., 2013, Kuo et al., 2021, Fotiou et al., 2022, Xia et al., 2024). In other work, selection is relevance-driven. SMART shares an experience only if inter-cell interference exceeds $I_{\min}\approx-110\,\mathrm{dBm}$ (Dahal et al., 27 Jan 2025). KVComm selects the top-$r$ fraction of layers or KV pairs using attention importance scores weighted by a Gaussian prior (Shi et al., 2 Oct 2025). OptimShare selects which attributes remain clear and which are perturbed, and then searches over $(\epsilon,\delta)$ pairs to maximize utility under privacy constraints (Chamikara et al., 2023).

This suggests that selective data sharing is best understood as a control problem over disclosure granularity, authorization semantics, and release modality. A plausible implication is that different subfields instantiate the same abstraction at different layers: policy resolution over principals, cryptographic transformation over data, and relevance filtering over representations.

2. Access-control semantics and recipient selection

A large portion of the literature formulates selective sharing through explicit authorization semantics. PeerShare defines policies on the social graph $G=(U,E)$, including “Friends,” “Friends-of-friends,” and custom lists $L(u)$. The server resolves a policy into a concrete recipient set $R\subseteq U$, and for a request $\mathrm{DOWNLOAD}(u)$ returns

$$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$

The notable property is that high-level social policies are reduced to explicit server-side recipient sets, which are then enforced on download (Nagy et al., 2013).

3LSAA implements a different access-control semantics. Search is first limited by SSE trapdoors, including a subset-search extension in which a trapdoor

$PRF_c$0

restricts matching to a subset $PRF_c$1. Access is then refined by ABE policies $PRF_c$2 over attributes, and only users whose credentials satisfy $PRF_c$3 can proceed to local AES-key recovery (Kuo et al., 2021). The resulting authorization chain is conjunctive: keyword consent, subset authorization, attribute satisfaction, and possession of recovery trapdoors all matter.

SelectShare encodes rights in a Verifiable Credential whose credentialSubject.capabilities claim maps device identifiers to the exact fields a client may access. The proxy verifies that the requested $PRF_c$4 set is a subset of those capabilities, validates that the VC is signed, unexpired, not revoked, and audience-bound, and additionally checks proof-of-possession through DPoP (Fotiou et al., 2022). Here the selective object is not only the data record but the field set within a signed JSON object.

Programmable Dataflows generalizes authorization into a contract abstraction. A contract is

$PRF_c$5

and approval requires

$PRF_c$6

Execution then proceeds only if the precondition and postcondition succeed (Xia et al., 2024). This shifts selective sharing from static ACLs to negotiated dataflow authorization.

TD-DCSS makes the task itself the authorization primitive. A task $PRF_c$7 grants one-time decryption of exactly the subset of granules indexed by $PRF_c$8, binds consent to a specific service provider identifier $PRF_c$9, and supports immediate revocation via capsule update (Lyu et al., 2024). This design narrows selective release to “exactly the subset of granules” specified by the task, rather than to an entire file or record.

These systems reject the common “all or nothing” model in different ways. The common misconception is that selective sharing is merely a narrower ACL over the same payload. The cited schemes show that selection may instead be expressed over recipients, files, indices, fields, granules, or computations, and that the enforcement logic may be social, cryptographic, credential-based, or contractual (Fotiou et al., 2022, Lyu et al., 2024).

3. Cryptographic and system architectures

Selective data sharing is implemented through markedly different trust and cryptographic assumptions. PeerShare does not introduce “a second layer of encryption or per-recipient key wrapping.” Instead, it relies on transport-level protection: all client–server communication runs over TLS, certificate pinning rejects non-matching server certificates, OAuth2 bearer tokens are validated against Facebook’s Graph API, and the trusted central server resolves policies and enforces access (Nagy et al., 2013). The system therefore provides selective dissemination through authenticated transport and server-side policy resolution, not end-to-end cryptographic fan-out.

3LSAA adopts the opposite stance for an untrusted environment. Its three layers are Searchable Symmetric Encryption, Attribute-Based Encryption, and local AES key recovery. At the SSE layer, the server can test whether

$I_{\min}\approx-110\,\mathrm{dBm}$0

without learning the keyword. At the ABE layer, a user’s credentials are checked against the policy ciphertext. At the AES layer, the owner masks the file key as

$I_{\min}\approx-110\,\mathrm{dBm}$1

and authorized users recover $I_{\min}\approx-110\,\mathrm{dBm}$2 locally using owner- and AA-issued trapdoors (Kuo et al., 2021). The server sees blinded group elements and ciphertexts, and the owner retains actual access control.

The MCPS scheme uses a different decomposition: selective encryption plus fragmentation and dispersion. The file body $I_{\min}\approx-110\,\mathrm{dBm}$3 is partitioned into 256-bit blocks $I_{\min}\approx-110\,\mathrm{dBm}$4, each block is split into eight 32-bit fragments, one fragment $I_{\min}\approx-110\,\mathrm{dBm}$5 is selected pseudorandomly using $I_{\min}\approx-110\,\mathrm{dBm}$6, and the remaining 224 bits are masked by

$I_{\min}\approx-110\,\mathrm{dBm}$7

The aggregated private fragment $I_{\min}\approx-110\,\mathrm{dBm}$8 is then AES-encrypted and kept locally or on a higher-trust cloud, while $I_{\min}\approx-110\,\mathrm{dBm}$9 is uploaded to the public cloud (Qiu et al., 2019). The security claim is specifically that even with $r$0 and full $r$1, reconstruction still requires the missing $r$2.

TD-DCSS places a CP-ABE policy and an integrity tag inside a tamper-resistant capsule. The capsule

$r$3

is indexed by $r$4, and tamper resistance is checked through

$r$5

The same framework binds task tokens to $r$6, provides one-time download and revocation tokens, and updates only parts of the capsule during revocation (Lyu et al., 2024).

Data Station uses a systems-oriented architecture rather than a purely cryptographic one. Data Elements are encrypted at rest under per-owner symmetric keys; a FUSE-based Interceptor mediates file-system access; execution runs inside AMD SEV-SNP; remote attestation verifies the enclave image; and policies of the form $r$7 are checked before decryption and mounting (Xia et al., 2023). In near-zero-trust mode, confidentiality depends on “DE decryption only inside enclave” and attested code identity, while auditability derives from an append-only log signed by the enclave key.

These architectures illustrate an important divide. Some systems treat selective sharing as trusted mediation by a central service (Nagy et al., 2013). Others assume an honest-but-curious or untrusted server and therefore hide search, policy, or key material cryptographically (Kuo et al., 2021, Qiu et al., 2019). Data Station occupies an intermediate position, moving trust into attested hardware and audited delegated execution (Xia et al., 2023).

4. Minimal release units, selective disclosure, and revocation

A central question is the unit at which disclosure becomes selective. The literature shows that this unit can be much smaller than a file or database row.

In the MCPS method, the selective unit is the per-block private fragment $r$8, which is only $r$9 bits out of each $(\epsilon,\delta)$0-bit block, i.e., $(\epsilon,\delta)$1 of the block; the public fragment $(\epsilon,\delta)$2 is $(\epsilon,\delta)$3 bits, i.e., $(\epsilon,\delta)$4 (Qiu et al., 2019). The public cloud stores the large fragment and the trusted smartphone controls release of the small fragment. This architecture realizes selective sharing by asymmetrically partitioning reconstructive information.

In TD-DCSS, the “minimal” unit is explicit: the Data Capsule is “the minimal unit of storage/share,” while the task $(\epsilon,\delta)$5 allows decryption of “exactly the subset of granules with indices $(\epsilon,\delta)$6” (Lyu et al., 2024). The scheme also makes revocation immediate. Upon receiving $(\epsilon,\delta)$7, the cloud server updates $(\epsilon,\delta)$8, modifies $(\epsilon,\delta)$9, recomputes $G=(U,E)$0, and outputs the updated capsule in $G=(U,E)$1 time and with constant-size tokens (Lyu et al., 2024). This is a notably strong notion of post-issuance control.

SelectShare makes the selective unit a revealed subset of signed messages. The transcoder signs canonicalized JSON with a BBS+ multi-message signature

$G=(U,E)$2

and the proxy later returns only the framed JSON plus a non-interactive proof that the revealed subset is correctly derived from the original signed item (Fotiou et al., 2022). The client or third party verifies two pairing-based relations, so integrity and authenticity apply to the subset without exposing the hidden fields.

PeerShare supports per-item policies rather than field-level proof systems. Each AppData item carries a value, type, owner, and sharing policy, and only the creating app may modify or delete that item, as identified by package name and developer key (Nagy et al., 2013). This is selective sharing at the item granularity.

The difference among these schemes is not only cryptographic style but release semantics. SelectShare preserves integrity of a subset of a signed object (Fotiou et al., 2022). TD-DCSS couples minimal-unit release with one-time authorization and immediate revocation (Lyu et al., 2024). MCPS separates reconstruction-critical and non-critical fragments (Qiu et al., 2019). PeerShare scopes entire application data items to social recipient sets (Nagy et al., 2013). This suggests that “selective” should not be restricted to access predicates alone; it also concerns the granularity at which authenticity, revocability, and reconstructability are preserved.

5. Delegated computation, contracts, and owner-controlled disclosure

A major branch of the literature avoids raw-data release altogether and instead makes computation the selectively shareable object. Data Station is explicit on this point: “Data owners share data with the escrow knowing it will not be released without their consent. Data users delegate their computation to the escrow” (Xia et al., 2023). The Gatekeeper forms an intent set $G=(U,E)$3, the Policy Broker checks whether each required triple is allowed, and only then are Data Elements mounted and decrypted inside the enclave (Xia et al., 2023). If the function is data-blind and accesses enclave-mode data, resulting Data Elements can be placed in a staging area until explicit policies are granted.

Programmable Dataflows abstracts this workflow into a formal data-sharing model. A state is $G=(U,E)$4, a dataflow is a transition $G=(U,E)$5, and contracts are used so that agents can “communicate the intent of a dataflow and evaluate its consequences, before the dataflow takes place” (Xia et al., 2024). The enforcement model is triphasic: precondition check, compute, postcondition check. This makes selective sharing programmable at the application level.

Two optimizations in Programmable Dataflows are directly relevant to selective sharing overhead. Intermediate outputs with empty destination sets can be cached and reused if a later dataflow has the same provenance, and illegal access can be short-circuited by aborting execution when $G=(U,E)$6 (Xia et al., 2024). In the ad-matching microbenchmark, caching reduces a 5-run workload from $G=(U,E)$7 to $G=(U,E)$8 at $G=(U,E)$9 M records per site, a $L(u)$0 speedup. In the fraud-detection microbenchmark, short-circuiting reduces runtime from $L(u)$1 to $L(u)$2 at $L(u)$3 GB, a $L(u)$4 speedup (Xia et al., 2024).

Data Station likewise emphasizes that controlled release can be practical. In its evaluation, owner-side registration and policy creation remain below $L(u)$5 each even in near-zero-trust mode, and the policy-listing overhead is “~0.1 s for 500 DEs” and “~12 s for 5 000 DEs×100 functions” (Xia et al., 2023). The enclave overhead on TPC-H $L(u)$6 GB is “5%–20%” for most queries, with the worst write-heavy query at “~35%” (Xia et al., 2023).

This branch of the literature reframes selective data sharing as selective dataflow authorization. A plausible implication is that it partially dissolves the tension between sharing and non-sharing: parties can pool data value without transferring raw datasets to each other, provided they accept an escrow or enclave-based execution model (Xia et al., 2023, Xia et al., 2024).

6. Efficiency trade-offs, domain-specific instantiations, and broader interpretations

Selective data sharing is often motivated by an overhead–utility trade-off rather than by confidentiality alone. SMART demonstrates this in decentralized multi-agent reinforcement learning for inter-cell interference management. Each base station shares experiences only when

$L(u)$7

so only “high-interference” samples are transmitted (Dahal et al., 27 Jan 2025). In the reported $L(u)$8-cell, $L(u)$9-UE per cell mmWave simulations, SMART shared only “≈25% of all experiences” and achieved network sum-rate “within 1–2% of the full-sharing (‘Share All’) and CTDE baselines,” with communication overhead “roughly 25% versus 100% under full sharing” (Dahal et al., 27 Jan 2025). Here selection is by relevance, not by privacy policy.

KVComm applies a similar principle to multi-LLM communication. A sender computes KV caches for a long context, scores layers using average attention

$R\subseteq U$0

weights them with a Gaussian prior $R\subseteq U$1, and selects the top-$R\subseteq U$2 layers according to $R\subseteq U$3 (Shi et al., 2 Oct 2025). The reported transmission ratios are $R\subseteq U$4, and with only “30% of layers’ KV pairs” KVComm “outperforms all prior communication schemes,” while “50–70%” can “match or even slightly surpass Skyline on many datasets” and reduce FLOPs by “2.5–6×” compared to Skyline (Shi et al., 2 Oct 2025). This is selective sharing over internal model state.

OptimShare addresses a different trade-off: controlled partially perturbed release for tabular data. The released table is

$R\subseteq U$5

where quasi-identifiers remain clear, sensitive attributes are perturbed by a DP mechanism, and $R\subseteq U$6 are selected using a fuzzy model driven by the Personal Information Factor (Chamikara et al., 2023). The paper reports, for example, on NHDS $R\subseteq U$7 an average utility of $R\subseteq U$8 and average effectiveness of $R\subseteq U$9, and on LRDS $\mathrm{DOWNLOAD}(u)$0 an average utility of $\mathrm{DOWNLOAD}(u)$1 and average effectiveness of $\mathrm{DOWNLOAD}(u)$2 (Chamikara et al., 2023). In the NHDS case study, OptimShare keeps $\mathrm{DOWNLOAD}(u)$3 at $\mathrm{DOWNLOAD}(u)$4 versus $\mathrm{DOWNLOAD}(u)$5 in the original, whereas DP-WGAN gives $\mathrm{DOWNLOAD}(u)$6 (Chamikara et al., 2023). The selective act here is attribute release in the clear versus under perturbation.

The economic literature studies selection at the level of consumer information. In the Hotelling model of Pareto-Improving Data-Sharing, a mechanism $\mathrm{DOWNLOAD}(u)$7 shares exact locations only for subsets $\mathrm{DOWNLOAD}(u)$8 and $\mathrm{DOWNLOAD}(u)$9. In the evenly split case, the consumer-optimal mechanism sets

$$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$0

keeps uniform prices at $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$1, and raises welfare from $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$2 to $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$3 (Gradwohl et al., 2022). The firm-optimal Pareto-improving mechanism shares narrower intervals and raises each firm’s profit from $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$4 to $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$5 while keeping consumer welfare weakly above $$\{ d \mid d.owner \neq u \wedge u \in R(d) \} \cup \{ d \mid d.owner = u \}.$$6 (Gradwohl et al., 2022). This is a selective sharing mechanism in a mechanism-design sense rather than a cryptographic one.

Taken together, these results show that selective data sharing is not a single protocol family. It includes secure dissemination among social contacts, owner-sovereign encrypted search and access control, user-centric fragmentation, task-bound capsules, VC/ZKP-based subset disclosure, enclave-mediated delegated computation, programmable dataflow contracts, relevance-filtered exchange in decentralized learning, and even partial information-sharing rules in markets (Nagy et al., 2013, Kuo et al., 2021, Qiu et al., 2019, Lyu et al., 2024, Fotiou et al., 2022, Xia et al., 2023, Xia et al., 2024, Dahal et al., 27 Jan 2025, Shi et al., 2 Oct 2025, Gradwohl et al., 2022). A plausible implication is that the field is converging on a common systems principle: transmit, reveal, or compute on only the smallest subset that is necessary for the intended objective, and encode the admissible subset in a verifiable control plane.