Dependent Types For Cryptography
Implementations
Paulo Silva

Manuel Barbosa

HASLab, Departamento de Informática
Universidade do Minho
Portugal

June 14, 2011
Motivation

Cryptographic software demands high-quality
implementations
The CAO language was developed close to cryptographic
standards making the implementation easier and more
reliable
This language is strongly typed with explicit type sizes
Improves safety but makes it less general and usable
Proposed solution: dependent types ⇒ CALF language
CAO Language
Small and simple domain specific language with imperative
flavour
Geared toward the automatic production of highly efficient
target code subject to security-aware optimizations
Type system supports cryptography types such as bit
strings, matrices and field extensions
CAO has a complete formalization of its:
Syntax
Semantics
Type system

We have proved that CAO type system is sound, i.e.,
“well-typed programs do not go wrong”
A fully functional CAO interpreter is also available
CAO Example
AES fragment

typedef GF2 := mod[ 2 ];
typedef GF2N :=
mod[ GF2<X> / X**8 + X**4 + X**3 + X + 1 ];
typedef S
:= matrix[4,4] of GF2N;
def mix : matrix[4,4] of GF2N
{[X], [X+1],[1], [1],
[1], [X], [X+1],[1],
[1], [1], [X], [X+1],
[X+1],[1], [1], [X]};

:=

def MixColumns( s : S ) : S {
def r : S;
seq i := 0 to 3 {
r[0..3,i] := mix * s[0..3,i]; }
return r; }
Limitations of CAO

In CAO all type sizes have to be statically determined
In the previous example, the MixColumns function only
works with 4 × 4 matrices
We would like to allow parametrisation of these sizes. For
instance:
typedef S<(n : int)> := matrix[n,n] of GF2N;
def MixColumns<(n : int)>( s : S<(n)> ) : S<(n)> {
def r : S<(n)>;
seq i := 0 to n-1 {
r[0..n,i] := mix * s[0..n,i]; }
return r; }
Dependent types
A dependent type depends on a value belonging to the
realm of program expressions
Can be seen as families of types indexed by values
In polymorphism, the type depends on another type
parameter, e.g.,
∀ α ∈ types . Vector of α
leading to vectors of integers, vectors of booleans, etc.
Using dependent types, the type depends on a value, e.g.,
Π n : Int . Vector[n]
leading to vectors of length 5, vectors of length 13, etc.
Dependent types
Dependent types allow for specification of program
properties in types, reducing verification of correctness to
type checking
Implementation and specification are kept synchronized
However, type checking of programs using full-fledge
dependent types is not decidable and cannot be done
automatically
To overcome this problem, is is necessary to limit their
expressive power reducing the amount of verifiable
properties
Most existing work is theoretical or in the context of
functional languages
CALF Language

CALF is a higher-level extension of the CAO language,
additionally providing:
Dependent types
Higher-order polymorphic operators (map, fold, and
zip-with)
User-defined parametric data types
Explicit constant definitions
Module system (allowing module instantiation)
CALF Language

The CALF compiler translates CALF source code to CAO
CALF programs are like templates which can be
instantiated with concrete values, leading to multiple CAO
programs
Dependent types allow for verifying some important
properties, without requiring code annotations or deductive
tools, directly in the generic CALF code
For instance, this allows for detecting many out-of-bounds
accesses in vectors, matrices or bit strings
The translation guarantees the safety properties
Dependent types in CALF
CALF has three different kinds of variable-like identifiers:
Language variables
Constants
Index variables

All variable-like identifiers have to be explicitly declared
with their respective type (type inference may be
considered in the future)
Index variables allow the introduction of dependent types
These are variables which can be used, not only in type
declarations, but also in program expressions
In the scope of their declaration, they are treated as
constants
They can be instantiated with any value of their domain
type
Type Expression Evaluation and Type Equality

The implementation of dependent types poses two key
questions:
How to deal with type expressions which are not known at
compile time?
How to define equality, since we cannot rely on syntactic
equality any more?

CALF evaluation mechanism deals with type expressions
that either evaluate to a value or to an expression
depending solely on index variables
Type equality is defined in evaluated type expressions,
possibly generating additional constraints
Type Equality Decision

Two approaches are used to solve generated constraints to
decide equality:
Syntactic manipulation of the constraint expressions
A Satisfiability Modulo Theories (SMT) solver

In our approach, two index variables are equal if and only if
they have the same symbolic value
Some additional restrictions (not discussed here) are
imposed in order to guarantee a less complex
implementation while maintaining the expressive power
In practice, we often need unification and substitution
instead of equality
Safety conditions

Sometimes the constraints cannot be verified although the
program is correct
Given a set of constraints, we have three possible results:
The constraints are satisfied — The code is safe
The exists one value for which the constraints are not
satisfied — The code is not safe
It is not possible to decide if the constraints are satisfied —
Unknown case

In the last case, the result is set by the user: succeed,
issue a warning or fail
Translation from CALF to CAO

The translation requires two files:
CALF source file Definition of data types, constants and
function
Specification file Concrete instantiations for the global
index variables
When modules are used, the import declarations have to
be checked and processed accordingly
Translation from CALF to CAO

The process occurs in three phases:
1
2

3

The CALF source file is type checked
The specification file is type checked against the
information collected during the previous phase. A list of
substitutions is returned with the required instantiations.
This list of substitutions is used to generate the output CAO
source. This requires collecting all dependencies between
functions and types

Several instances of the same function or data type may
be generated
CALF Example
RSA fragment

typedef RSAPub<(m : int)>
:=
struct [ def encExp : int; ];
typedef RSAPrivShort<(m : int)> :=
struct [ def decExp : int; ];
def RSA<(n : int)>(k : RSAPub<(n)>, m : int ) : int {
def c : mod[n];
c := (mod[n]) m; c := c ** k.encExp;
return (int) c;
}
def RSAInvShort<(n : int)>
(k : RSAPrivShort<(n)>, c : int) : int {
def m : mod[n];
m := (mod[n]) c;
return (int) m;
}

m := m ** k.decExp;
CALF Example
RSA fragment

def const pq : int;
def const d : int;
def const e : int;
def x : int;
def y : int;
def myPub : RSAPub<(pq)>;
def myPriv : RSAPrivShort<(pq)>;
def Calc() : void {
myPub.encExp := e;
y := RSA<(pq)>(myPub,x);
}
CALF Example
Specification file

def const pq : int := 35;
def const d : int := 11;
def const e : int := 11;
CALF Example
Generated CAO code

typedef RSAPub_35 := struct[def encExp_35 : int;];
def RSA_35(k : RSAPub_35, m : int) : int {
def c : mod[35];
c := (mod[35]) m;
c := c ** k.encExp_35;
return (int) c;
}
def myPub : RSAPub_35;
def x : int;
def y : int;
def Calc() : void {
myPub.encExp_35 := 11;
y := RSA_35(myPub, x);
}
The Overall Picture
Ongoing Work

Introducing explicit constraints in index variables (very
important for practical usage)
Improving the generation and solving of constraints in
iterative statements
Improving the module system (object oriented?)
Publication of results

Dependent Types for Cryptography Implementations

  • 1.
    Dependent Types ForCryptography Implementations Paulo Silva Manuel Barbosa HASLab, Departamento de Informática Universidade do Minho Portugal June 14, 2011
  • 2.
    Motivation Cryptographic software demandshigh-quality implementations The CAO language was developed close to cryptographic standards making the implementation easier and more reliable This language is strongly typed with explicit type sizes Improves safety but makes it less general and usable Proposed solution: dependent types ⇒ CALF language
  • 3.
    CAO Language Small andsimple domain specific language with imperative flavour Geared toward the automatic production of highly efficient target code subject to security-aware optimizations Type system supports cryptography types such as bit strings, matrices and field extensions CAO has a complete formalization of its: Syntax Semantics Type system We have proved that CAO type system is sound, i.e., “well-typed programs do not go wrong” A fully functional CAO interpreter is also available
  • 4.
    CAO Example AES fragment typedefGF2 := mod[ 2 ]; typedef GF2N := mod[ GF2<X> / X**8 + X**4 + X**3 + X + 1 ]; typedef S := matrix[4,4] of GF2N; def mix : matrix[4,4] of GF2N {[X], [X+1],[1], [1], [1], [X], [X+1],[1], [1], [1], [X], [X+1], [X+1],[1], [1], [X]}; := def MixColumns( s : S ) : S { def r : S; seq i := 0 to 3 { r[0..3,i] := mix * s[0..3,i]; } return r; }
  • 5.
    Limitations of CAO InCAO all type sizes have to be statically determined In the previous example, the MixColumns function only works with 4 × 4 matrices We would like to allow parametrisation of these sizes. For instance: typedef S<(n : int)> := matrix[n,n] of GF2N; def MixColumns<(n : int)>( s : S<(n)> ) : S<(n)> { def r : S<(n)>; seq i := 0 to n-1 { r[0..n,i] := mix * s[0..n,i]; } return r; }
  • 6.
    Dependent types A dependenttype depends on a value belonging to the realm of program expressions Can be seen as families of types indexed by values In polymorphism, the type depends on another type parameter, e.g., ∀ α ∈ types . Vector of α leading to vectors of integers, vectors of booleans, etc. Using dependent types, the type depends on a value, e.g., Π n : Int . Vector[n] leading to vectors of length 5, vectors of length 13, etc.
  • 7.
    Dependent types Dependent typesallow for specification of program properties in types, reducing verification of correctness to type checking Implementation and specification are kept synchronized However, type checking of programs using full-fledge dependent types is not decidable and cannot be done automatically To overcome this problem, is is necessary to limit their expressive power reducing the amount of verifiable properties Most existing work is theoretical or in the context of functional languages
  • 8.
    CALF Language CALF isa higher-level extension of the CAO language, additionally providing: Dependent types Higher-order polymorphic operators (map, fold, and zip-with) User-defined parametric data types Explicit constant definitions Module system (allowing module instantiation)
  • 9.
    CALF Language The CALFcompiler translates CALF source code to CAO CALF programs are like templates which can be instantiated with concrete values, leading to multiple CAO programs Dependent types allow for verifying some important properties, without requiring code annotations or deductive tools, directly in the generic CALF code For instance, this allows for detecting many out-of-bounds accesses in vectors, matrices or bit strings The translation guarantees the safety properties
  • 10.
    Dependent types inCALF CALF has three different kinds of variable-like identifiers: Language variables Constants Index variables All variable-like identifiers have to be explicitly declared with their respective type (type inference may be considered in the future) Index variables allow the introduction of dependent types These are variables which can be used, not only in type declarations, but also in program expressions In the scope of their declaration, they are treated as constants They can be instantiated with any value of their domain type
  • 11.
    Type Expression Evaluationand Type Equality The implementation of dependent types poses two key questions: How to deal with type expressions which are not known at compile time? How to define equality, since we cannot rely on syntactic equality any more? CALF evaluation mechanism deals with type expressions that either evaluate to a value or to an expression depending solely on index variables Type equality is defined in evaluated type expressions, possibly generating additional constraints
  • 12.
    Type Equality Decision Twoapproaches are used to solve generated constraints to decide equality: Syntactic manipulation of the constraint expressions A Satisfiability Modulo Theories (SMT) solver In our approach, two index variables are equal if and only if they have the same symbolic value Some additional restrictions (not discussed here) are imposed in order to guarantee a less complex implementation while maintaining the expressive power In practice, we often need unification and substitution instead of equality
  • 13.
    Safety conditions Sometimes theconstraints cannot be verified although the program is correct Given a set of constraints, we have three possible results: The constraints are satisfied — The code is safe The exists one value for which the constraints are not satisfied — The code is not safe It is not possible to decide if the constraints are satisfied — Unknown case In the last case, the result is set by the user: succeed, issue a warning or fail
  • 14.
    Translation from CALFto CAO The translation requires two files: CALF source file Definition of data types, constants and function Specification file Concrete instantiations for the global index variables When modules are used, the import declarations have to be checked and processed accordingly
  • 15.
    Translation from CALFto CAO The process occurs in three phases: 1 2 3 The CALF source file is type checked The specification file is type checked against the information collected during the previous phase. A list of substitutions is returned with the required instantiations. This list of substitutions is used to generate the output CAO source. This requires collecting all dependencies between functions and types Several instances of the same function or data type may be generated
  • 16.
    CALF Example RSA fragment typedefRSAPub<(m : int)> := struct [ def encExp : int; ]; typedef RSAPrivShort<(m : int)> := struct [ def decExp : int; ]; def RSA<(n : int)>(k : RSAPub<(n)>, m : int ) : int { def c : mod[n]; c := (mod[n]) m; c := c ** k.encExp; return (int) c; } def RSAInvShort<(n : int)> (k : RSAPrivShort<(n)>, c : int) : int { def m : mod[n]; m := (mod[n]) c; return (int) m; } m := m ** k.decExp;
  • 17.
    CALF Example RSA fragment defconst pq : int; def const d : int; def const e : int; def x : int; def y : int; def myPub : RSAPub<(pq)>; def myPriv : RSAPrivShort<(pq)>; def Calc() : void { myPub.encExp := e; y := RSA<(pq)>(myPub,x); }
  • 18.
    CALF Example Specification file defconst pq : int := 35; def const d : int := 11; def const e : int := 11;
  • 19.
    CALF Example Generated CAOcode typedef RSAPub_35 := struct[def encExp_35 : int;]; def RSA_35(k : RSAPub_35, m : int) : int { def c : mod[35]; c := (mod[35]) m; c := c ** k.encExp_35; return (int) c; } def myPub : RSAPub_35; def x : int; def y : int; def Calc() : void { myPub.encExp_35 := 11; y := RSA_35(myPub, x); }
  • 20.
  • 21.
    Ongoing Work Introducing explicitconstraints in index variables (very important for practical usage) Improving the generation and solving of constraints in iterative statements Improving the module system (object oriented?) Publication of results