Page 35 - Invited Paper Session (IPS) - Volume 1

IPS35 Dinov I.D. et al.
S(f ), such that ⱯD1, D2 that differ on at most one element IIf (D1) - f (D2)
II1 ≤ S(f ). There are many differentially private algorithms, e.g., random
forests, decision trees, k-means clustering, etc. For instance, if f : D = DB
m
→ R , the algorithm outputting f (D ) + (ղ1, ղ2, … , ղm), with ղί € Laplace
, Ɐ ί is ℇ-differently private.

2.2 Fully-Homomorphic Encryption (FHE)
FHE security is based on preprocessing the data by encryption to allow
subsequent program execution and data-driven inference using the encrypted
information (Gentry 2009). As a result, the process outputs are encrypted and
their interpretation requires ability to decrypt the information following the
data analytics. It represents an elegant and powerful mathematical framework
for bijective (encoding/decoding) processing and analytics. Albeit, it is very
fast, FHE has some limitations, e.g., deriving the f ′ – commutative analytic
evaluators – is never a trivial task and requires close cooperation between data
governor and data user. Figure 2 shows schematically the process of data
analytics using fully-homomorphic encryption.

Figure 2: Data analytics via fully-homomorphic encryption.

2.3 DataSifter Statistical Obfuscation
The process of data-masking using statistical obfuscation is the core of the
DataSifter technique. It combines artificial random missingness with partial
information alterations using data swapping within subjects’ neighbourhoods.
These operations have minimal impact on the joint distribution of the
obfuscated (sifted) output data as the controlled rate of missingness is
introduced completely at random and nearest neighbourhoods tend to have
consistent distributions. The DataSifter algorithm preserves the exact data

24 | I S I W S C 2 0 1 9