Inferring Semantic Mapping Between Policies and Code: The Clue is in the Language

Abstract

A common misstep in the development of security and privacy solutions is the failure to keep the demands resulting from high-level policies in line with the actual implementation that is supposed to operationalize those policies. This is especially problematic in the domain of social networks, where software typically predates policies and then evolves alongside its user base and any changes in policies that arise from their interactions with (and the demands that they place on) the system. Our contribution targets this specific problem, drawing together the assurances actually presented to users in the form of policies and the large codebases with which developers work. We demonstrate that a mapping between policies and code can be inferred from the semantics of the natural language. These semantics manifest not only in the policy statements but also coding conventions. Our technique, implemented in a tool (CASTOR), can infer semantic mappings with F1 accuracy of 70 % and 78 % for two social networks, Diaspora and Friendica respectively – as compared with a ground truth mapping established through manual examination of the policies and code.

Appendices

A Implementation: CASTOR

We have implemented our technique in a tool called CASTOR. Figure 4 illustrates the architecture of CASTOR. CASTOR accepts as inputs policy statements and source code; and outputs a set of semantic mappings between policy statements and functions. Briefly, CASTOR works on the input as follows:

Policy Engine: CASTOR’s policy engine is composed of a parser and a statement analyser which transforms the natural language policy into an intermediate representation (as described in Sect. 3.2). This intermediate representation maintains the relevant policy primitives of a statement, namely action (verbs) and data (nouns).

Code Engine: CASTOR’s code engine is composed of a minimal recursive-descent parser that extracts a function’s name, associated class and parameters, along with information identifying the source file and line number where the function can be found. This is inline with our source model construction in Sect. 3.3.

Fig. 4.

CASTOR’s architecture.

Mapping Engine: CASTOR’s mapping engine infers the mapping between the privacy policy $\mathcal {PP}$ and source code functions $\mathcal {F}$ using its inbuilt WordNet corpora and classifier. The output of this engine is a set of semantic mappings between policy statement(s) and functions.

B Formulae

Recall (TPR) $= \frac{tp}{tp + fn}$; False-Positive Rate (FPR) $= \frac{fp}{fp + tn}$; Precision (PPV) $= \frac{tp}{tp + fp}$; and F1 $= 2 \cdot \frac{Precision \cdot Recall}{Precision + Recall}$.