|
| Method | Encode | Randomness | Asymptotic bound error | Candidate | Communication cost | Computation
cost | Pros and cons |
|
-RR [35]
GRR [36] | Direct | Local | | Known | P: Θ
S: Θ | P: O (1)
S: O | Pros: no encoding, predigest the process; lower candidate size can achieve higher utility; cons: low utility in low privacy regime |
|
| O-RR [35] | Unary (bloom filter) | Local | | Unknown | P: O (h)
S: O (nh) | P: O
S: Linear regression | Pros: open candidate; cons: low utility in low privacy regime, high computation cost due to regression |
|
| RAPPOR [7] | Unary (bloom filter) | Local | | Known | P: O (h)
S: O (nh) | P: O
S: LASSO and linear regression | Pros: lower error, lower storage cost, support big candidate; cons: consider Bloom filter parameter settings, high computation cost due to regression |
|
| k-RAPPOR (basic one-time) [7] | Unary | Local | | Known | P: Θ (k)
S: O (nk) | P: O
S: | Pros: lower error, lower storage overhead, simpler and faster implement; cons: consider parameter settings of Bloom filter |
|
| OUE [36] | Unary | Local | | Known | P: Θ (k)
S: O (nk) | P: O
S: | Pros: lower error, lower storage cost, lower computation cost and easier to implement; cons: larger candidate lead to higher communication cost |
|
| O-RAPPOR [35] | Unary (bloom filter) | Local | | Unknown | P: Θ (h)
S: O (nh) | P: O
S: linear regression | Pros: open candidate, higher utility, lower storage overhead; cons: need consider parameter settings of bloom filter |
|
| k-Subset [41, 42] | Direct | Local | | Known | P: Θ (k)
S: O (nk) | P: O
S: | Pros: better sample complexity and higher utility; cons: higher communication and computation cost due to set output |
|
| RMP (SHist) [37] | Binary | Public (shared matrix) | | Known | P: O (1)
S: O (n) | P: O
S: O | Pros: lower communication cost; cons: unstable query accuracy due to the noise from RMP matric |
|
| HRR [10, 38] | Binary | Public (shared matrix) | | Known | P: O (1)
S: O (n) | P: O
S: O | Pros: lower communication cost; cons: unable query accuracy due to the noise from RMP matric |
|
| BLH [36] | Binary | Local and public | | Known | P: O (1)
S: Θ (log(n)) | P: O
S: O | Same with the RMP method |
|
| OLH [36] | Binary | Local and public | | Unknown | P: O (1)
S: Θ (log(n)) | P: O (k)
S: O | Pros: higher utility in the setting big candidate size, lower communication cost; cons: unstable accuracy due to the noise from RMP matric |
|
| HR [39] | Binary | Local | | Known | P: O (log (k))
S: (O (nlog (k)) | P: O
S: O | Pros: obtain efficient computation complexity due to Fast Walsh–Hadamard transform; cons: unstable accuracy due to the noise from encoding |
|
