Advertisement Internship at Microsoft Research? 12 week research projects, undertaken at MSR Cambridge, typically by grad students mid-way through their PhD. Goal: complete and publish research project with an MSR researcher: K..Bhargavan, C. Fournet, A. Gordon, and R. Pucella, TulaFale: A security tool for web services, FMCO 2003 C. Fournet, A. Gordon, and S. Maffeis A type discipline for authorization policies, ESOP 2005 Applications for Summer 2006 are due by end February 2006 http://research.microsoft.com/ aboutmsr/jobs/internships/cambridge.aspx 1 From Typed Process Calculi to Source-Based Security Andy Gordon (MSR) Based on joint work with Cédric Fournet (MSR), Alan Jeffrey (DePaul and Bell Labs), and Sergio Maffeis (Imperial) SAS 2005, London September 7-9, 2005 Background Process calculi are an effective setting for modelling security protocols and specifying their properties Lowe (1995) used CSP to find his famous attack on the Needham-Schroeder public key protocol (1978) The spi calculus (AG97) began a line of work in which many protocols have been expressed and analyzed within pi calculi Security types allow the typechecker to prove various security properties automatically Syntax-driven typing rules can be checked efficiently, with no state space exploration Properties of arbitrarily many sessions and principals proved relative to arbitrary Dolev-Yao opponent Inevitably incomplete as D-Y problem undecidable (DLMS99) 3 An Authentication Example Suppose A and B are principals sharing a symmetric-key KAB The following should ensure B gets a fresh message from A Begin Assertion Message 1 Message 2 End Assertion A begins B A: A B: B ends Sent(A,B,msg) nonce A, {msg,nonce}KAB Sent(A,B,msg) We specify the authentication of the message via assertions: each end is to have distinct, preceding begin with same label Attacks (replays, impersonations) show up as violations of these assertions By assigning KAB the following type, we can check the protocol: Key (msg:T, Nonce [Sent(A,B,msg)] ) http://www.cryptyc.org 4 Authentication; full trust 2001 Secrecy; full trust Gordon/Jeffrey Typing correspondence assertions for communication protocols (MFPS 2001) 1999 Abadi Secrecy by typing in security protocols (JACM 1999) 2001 2001 Abadi/Blanchet Secrecy types for asymmetric communication (FOSSACS 2001) 2000 2002 2002 Abadi/Blanchet Analyzing security protocols with secrecy types and logic programs (POPL 2002) Applications2002 Lashari A polymorphic type and effect system for an object oriented language to typecheck cryptographic protocols (Masters, DePaul) 2004 Bugliesi/Focardi/Maffei Compositional analysis of authentication protocols (ESOP 2004) 2001 2002 2003 Type inference Other work on security in pi includes: Bodei/Degano/Nielson/Nielson Berger/Honda/Yoshida Gordon/Jeffrey Types and effects for asymmetic cryptographic protocols (CSFW 2002) 2003 Gordon/Pucella Validating a web services security abstraction by typing (XML Security 2003) 2005 2004 Fournet/Gordon/Maffeis A type discipline for authorization policies (ESOP 2005) 2005 2005 Focardi/Maffei/Placella Inferring authentication tags (WITS 2005) Gordon/Jeffrey Authenticity by typing for security protocols (CSFW 2001) 2005 Gordon/Jeffrey Secrecy despite compromise (CONCUR 2005) 2005 Haack/Jeffrey Timed spi-calculus with types (CONCUR 2005) Authorization; timing; partial trust This Talk Two new developments Checking authorization (is this request allowed?) as well as authentication (who sent this request?) Allowing a realistic threat model in which some trusted hosts become compromised over time A type discipline for authorization policies (With C. Fournet and S. Maffeis. ESOP'05) Secrecy despite compromise: types, cryptography, and the pi-calculus. (With A. Jeffrey. CONCUR'05) A useful idea in both is the use of inert processes to record events and to express security properties 6 A Type Discipline for Authorization Joint with C. Fournet and S. Maffeis Motivations Authorization policies prescribe conditions that must be satisfied before performing any privileged action In practice, policies often only formalized in code In principle, Hard to extract, hard to reason about, hard to audit Tied to low-level authentication mechanisms Relationship of code to intended policy left informal Policies can be formalized in high-level languages (e.g. Datalog) separate from the implementation code Policies should be independent of enforcement mechanisms Conformance of an implementation should be verifiable Our initial motivations Difficulty of auditing use of Java-style stack inspection Authorization for web services 8 Our Approach We propose language-based mechanisms to express the intended policy of an implementation, and to verify conformance to the policy We use the authorization policy as a specification Our implementation language is a spi calculus As opposed to being directly executed The same policy supports alternative implementations But the approach would apply to higher-level languages We use types to verify that annotated code correctly implements a given authorization policy 9 Datalog for Authorization Datalog is a fragment of Prolog without negation, free variables and term constructors Many policy languages for trust or authorization are based on Datalog or related logics (SD3, Binder, Cassandra, SPKI, XrML, …) Realistic policies: Becker’s 375 rule formalization of NHS Electronic Health Record system in Cassandra (CSFW’04) We use Datalog for specificity, but our results hold for any monotonic logic closed under substitutions 10 Ex: Conference Reviewing Extensional database: known facts (closed literals) These generalize the events, such as Sent(A,B,msg), used in direct correspondence assertions to specify authentication Rules for deriving new facts Intensional database: facts derived from rules 11 Spi calculus with annotations Zero-bits, only to keep track of guarantees Security annotations 12 Authorization Properties Inert processes model events and properties A statement C models part of the authorization policy Specifically, a fact L models an authorization event An expectation expect L models an expected property The structural equivalence PP’ and reduction PP’ relations are much as usual There are no rules for these inert processes 13 Some Basic Examples Process P specifying a policy and two facts: A robustly safe process : A safe process : … and the robustly safe version : 14 Authorization by Typing Every ok value must be justified Every binding occurrence may add facts in E 15 Type System: Results Verification is efficient Structural type system Low complexity of logical resolution 16 Typing the Examples Process P specifying a policy and two facts: A safe process (by typing) : A robustly safe process (by typing) : 17 In the Full Version Two distributed implementations of a policy for conference management One where each delegation is registered online The other enables offline, signature based delegation with authorization decisions based on certificate chains 18 Summary We used inert processes to annotate programs with expected authorization properties Extends work to typecheck direct correspondences Woo and Lam’s direct correspondences are derivable Much prior work on logics for authorization “At this point Report(U,ID,R) will be derivable” Goal: check code annotations against explicit logical policy Ours is amongst the first to relate such logics to code and to use DY approach to model untrusted parts of system Limitations: Like many systems, no support for revocation Interpreter + typechecker, but no direct implementation Principals completely distrusted or completely trusted... 19 Secrecy Despite Compromise Joint work with A. Jeffrey Motivation Our opponent model has assumed a fixed partition Trusted insiders versus distrusted outsiders Real situations are more complex Machines become compromised Trusted users turn out to be untrustworthy How can a type system handle partial compromise of a dynamically changing population of principals? We approach this question from a simpler setting than spi, Odersky’s polarized pi calculus Capabilities a? and a! for channel-based input and output 21 Security Levels Code annotated with security levels (or principals) Security ordering induced by arc processes Different regions may run on behalf of different levels Level annotation L attached to each output out a! M :: L Level represents the opponent Arc L1 L2 is itself an (inert) process Active (top-level) arcs in P induce a preorder P L1 L2 Least and greatest elements and Compound level (L1 , L2) has P (L1 , L2) Li for each i Security ordering represents compromise Let a level L be compromised iff L Hence L1 L2 means L1 is at risk of compromise by L2 So (L1, L2) is compromised if either L1 or L2 compromised 22 Security Hierarchies b a (a,b) any process b a (a,b) a any process a1 ... an G an+1 ... an+m !;new a;(G a | ;a ) 23 Conditional Secrecy We say M is public if it can be output at level We model secrecy invariants as inert processes: An expectation secret M amongst N is justified if every output of M is at a higher security level than N Read as “if M becomes public then N is compromised” The secret message M may include fresh names 24 A Basic Example Consider two processes at level L that exchange a fresh secret s on a private channel k We want a type system that: Checks secrecy of s while k is secret and L uncompromised Eventually allows k and s to be made public once L is compromised – an event modelled by the arc L A specific formal problem: verify robust safety of 25 Conditional Secrecy by Typing 26 In the Full Version Types ordered via a subtype relation Secrecy types are special case of (kinded) channels Kinds take the form {?L1,!L2} We can assert secrecy of channels, eg the k channel Type Ok{L1 L2} proves that L1 L2 Main rule: if Public(T) and Tainted(T’) then T <: T’ Allows security orderings to be communicated Type system reflects usage of pair types (split x:T, U) – first element extracted without checking (match x:T, U) – first element matched against known value Full form is (y x:T, U) where {split,match} and y is an existentially quantified lower bound on x used only in types 30 Typing a Crypto Protocol 31 Related Work Key or host compromise often modelled using events Bugliesi, Focardi, Maffei (FMSE’04) allow for compromised hosts in a type system for spi, but assume the set is known statically Types to govern data declassification are a Hot Topic Paulson (JCS 98): “oops” events mark key disclosure Myers and Liskov (TOSEM’00) DLM is one of the first system of security types to consider declassification, though at level of individual expressions, not types Several recent works (CSFW’05) on temporary modifications of a security ordering, akin to our L1 L2 processes Many studies of process calculi with security ordering Our use of an ordering to model runtime compromise is new 32 Summary, Conclusions We introduced a mutable security ordering to model a dynamic, partially compromised set of principals As with our authorization model, we rely on inert processes to describe events and expected properties There remains much promise in the area of process calculi with security types These two systems should combine fairly smoothly They should be applicable to an important open problem; how to check security properties of the actual source code of crypto protocols and the applications built on them 33 The End