Monte Carlo Analysis of Security Protocols: Needham-Schroeder Revisited Radu Grosu SUNY at Stony Brook Joint work with Xiaowan Huang, Scott Smolka, & Ping Yang June 8, 2004 -- DIMACS Workshop on Security Analysis of Protocols Talk Outline 1. LTL Model Checking 2. Monte Carlo Model Checking 3. Needham-Schroeder 4. Implementation & Results 5. Conclusions & Future Work Model Checking S | ? Is system S a model of formula φ? Model Checking • S is a nondeterministic/concurrent system. • is (in our case) an LTL (Linear Temporal Logic) formula. • Basic idea: intelligently explore S’s state space in attempt to establish S ⊨ . • Fly in the ointment: State Explosion! LTL Model Checking • An LTL formula is made up of atomic propositions p, boolean connectives , , and temporal modalities X (neXt) and U (Until). • Every LTL formula can be translated to a Büchi automaton whose language is set of infinite words satisfying . • Automata-theoretic approach: S ⊨ iff L(BS) L(B ) iff L(BS B ) Emptiness Checking • Checking non-emptiness is equivalent to finding an accepting cycle reachable from initial state (lasso). • Double Depth-First Search (DDFS) algorithm can be used to search for such cycles, and this can be done on-the-fly! sn sk+3 sk+2 sk+1 DFS2 s1 s2 s3 sk-2 sk-1 sk DFS1 Monte Carlo Model Checking (MC2) • Sample Space: lassos in BS B • Random variable Z : – Outcome = 0 if randomly chosen lasso accepting – Outcome = 1 otherwise • μZ = ∑ pi Zi (weighted mean) ~ of μ • Compute (ε,δ)-approx. Z Z Monte Carlo Model Checking (MC2) a b c d e L1 = abcb, L2 = abcdb, L3 = abcdea Pr[L1]= ½, Pr[L2]=¼, Pr[L3]=¼ μZ = ½ Monte Carlo Approximation • Problem: Compute the mean value μZ of a random variable Z distributed in [0,1] when an exact computation of μZ proves intractable. • Solution: Compute an (,)-approximation Z of Z: Pr[ Z (1 ) Z Z (1 )] 1 with error margin and confidence ratio . • Has been used to: approximate permanent of 0-1 valued matrices, volume of convex bodies, and, now, expectation that S ⊨ ! Original Solution [Karp, Luby & Madras: Journal of Algorithms 1989] • Compute Z as the mean value of N independent random variables (samples) identically distributed according to Z: Z ( Z1 ... Z N )/ N • Determine N using the Zero-One estimator theorem: N 4 ln(2 / )/ Z 2 • Problems: 1/ Z is unknown and 1/ 2 can be large. Stopping Rule Algorithm (SRA) [Dagum, Karp, Luby & Ross: SIAM J Comput 2000] • Innovation: computes correct N without using 1/ Z = 4 ln(2/) / 2; for (N=0, S=0; S≤; N++) S=S+ZN; Z = S/N; return ; Z • Theorem: Pr[ Z (1 ) Z Z (1 )] 1 E[N] ≤ 4 ln(2/) / μZ2; • Problem: 1/ 2 is in most interesting cases too large. Optimal Approx Algorithm (OOA) [Dagum, Karp, Luby & Ross: SIAM J Comput 2000] • Compute N using generalized Zero-One estimator: N 4 ln( 2 / )/ Z 4 ln( 2 / )/ Z 2 if σ Z2 Z otherwise • Apply sequential analysis (prediction/correction): 1. Assume 2 is small and compute ̂Z with SRA( , ) 2. Compute ̂ 2 using ̂Z and N 4 ln(2 / )/ ˆ 3. Use ̂ 2 to correct N and Z . Z • Expected number of samples is optimal to within a constant factor! Monte Carlo Model Checking Theorem: MC2 computes an (ε,δ)-approximation of μZ in expected time O(N∙D) and uses expected space O(D), where D is the recurrence diameter of B = BS B . Cf. DDFS which runs in O(2|S|+|φ|) time and space. Needham-Schroeder 1. A B : { Na, A } KB 2. B A : { Na, Nb } KA 3. A B : { Nb } KB Breaking & Fixing Needham-Shroeder • In 1997, Lowe discovered a replay attack that involves an intruder I masquerading as A in its communication with B. • As shown by Lowe, protocol is easily fixed by including identity of responder (B) in 2nd msg: 2´. B A : { B, Na, Nb } KA Implementation • Implemented DDFS and MC2 in jMocha model checker for synchronous systems specified using Reactive Modules. • Specified NS as a reactive module; all communications go through intruder. • Intruder obeys Dolev-Yao model: besides normal communications, can intercept, overhear, and fake messages. Experimental Results nonce (0..1) (0..4) (0..8) (0..20) (0..32) (0..36) (0..60) DDFS time entries 31 1 607 1 2527 2 11 24031 32 85279 46 18111 oom time 20 33 34 34 70 141 4200 MC2 exp 1 2 9 12 24 37 467 avg 12 29 30 30 30 30 30 Time and space requirements for DDFS and MC2 Experimental Results sat nonce 2915 (0..1) 2955 (0..4) 2969 (0..8) μ 2970 (0..20) 6288 (0..32) (0..36) 12975 (0..60) 194937 Z cntr 171 18 4 3 3 3 9 mu_Z 0.945 0.994 0.999 0.999 1 1 1 Variation of µ~Z for MC2 Related Approaches • NRL Protocol Analyzer [Meadows 96] • Spi-Calculus [Abadi Gordon 97] • FDR [Lowe 97] • The Strand Space Method [Guttman et al. 98] • Isabelle Theorem Prover [Paulson 98] • Backward Induction [Kurkowski Mackow 03] Conclusions • Applied Monte Carlo model checking to Needham-Schroeder. • Results indicate may be more effective than traditional approaches in discovering attacks. • Further experimentation required to draw definitive conclusions. • Other Future Work: Use BDDs to improve run time. Also, take samples in parallel! Monte Carlo Model Checking • Randomized algorithm for LTL model checking utilizing automata-theoretic approach. • Basic idea: Take N samples: sample = lasso = random walk through BS B ending in a cycle. • If accepting lasso (counter-example) found, return false. • Else return true with certain confidence.