what’s the scuttlebutt on secure gossip ?
DESCRIPTION
What’s the scuttlebutt on Secure Gossip ?. Robbert van Renesse. Secure Gossip or Gossip for Security?. Secure gossip: making sure probabilistic guarantees hold in the face of malicious behavior - PowerPoint PPT PresentationTRANSCRIPT
Secure Gossip orGossip for Security?
• Secure gossip: making sure probabilistic guarantees hold in the face of malicious behavior
• Gossip for security: exploit robustness of gossip for spreading important information (certificate revocation lists, security patches, …)
movement from former to latter?
Scale Security
• Gossip often said to scale well
• But in large settings more chance of bad behavior of participants– Spreading false rumors– Introducing too many rumors– Not forwarding good rumors– …
What have people worked on?
• Integrity– Preventing spreading of false gossip
without the use of digital signatures
• Availability– Making sure everybody receives all gossip
• Fairness– Everybody does its share of forwarding
8 years - 8 papers
• 1999 Malkhi, Mansour, Reiter– On diffusing updates in a Byzantine environment
• 2001 Malkhi, Reiter, Sella– Optimal unconditional information diffusion
• 2003 Minsky, Schneider– Tolerating malicious gossip
• 2003 Malkhi, Pavlov, Sella– Gossip with malicious parties
• 2004 Badishi, Keidar, Sasson– Exposing and eliminating vulnerabilites to DoS attacks in secure gossip-based
multicast
• 2006 Johansen, Allavena, Van Renesse– Fireflies: Scalable Support for Intrusion-Tolerant Network Overlays
• 2006 Haridasan, Van Renesse– Defense Against Intrusion in a Live Streaming Multicast System
• 2006 Li, Clement, Wong, Napper, Roy, Alvisi, Dahlin– BAR Gossip
Malkhi et al. take 1(Srikanth/Toueg)
• Integrity without digital signatures– Assume initially at least t +1 nodes have valid
gossip– A node does not forward gossip until it thinks it is
valid– A gossip becomes valid when it has been received
from at least t + 1 nodes
• Bad nodes cannot generate valid gossip• Slow…
Malkhi et al. take 2
– Initially t + 1 nodes have valid gossip– Nodes maintain set of source paths for each
gossip. Initially 1 empty path.– Gossip messages contain all paths.– On receipt of gossip from p, a node adds p to all
paths before adding to set of paths.– On receipt of t + 1 disjoint paths, gossip can be
delivered.
• Fast, but very large messages
Malkhi et al. take 3Minsky & Schneider
• Practical versions that only maintain a subset of paths
• Useful up to a dozen bad nodes or so
But there’s always public key signatures…
Keidar et al.
• Prevent application-level DoS attacks– Send digests on publicly known ports,
actual data on randomly generated ports– Randomly pick messages to exchange
from digests to prevent being overwhelmed by bogus messages
Van Renesse, Allavena, JohansenFOO gossip (Fireflies Ordered Rings)
• Large fraction of bad nodes– Generate a pseudo-random mesh that connects
all correct nodes with high probability– Gossip on mesh, using public key crypto for
integrity– Use persistent connections between nodes,
reducing need for full merges– Throttle state transfer for joins
• Moderate scalability, but can be fixed I think
Haridasan, Van RenesseSecureStream
• Video streaming with many bad hosts– Carefully flood video frames on pseudo-
random mesh using a notify/pull strategy– Prefer nodes that sent data to you over
nodes that don’t
• Deals pretty well with freeloaders
• Approach under development deals well with heterogeneity as well
Li, Alvisi et al.BAR gossip
• Elegant game-theoretic approach to dealing with rational and Byzantine nodes
(See RFC3092 for FOOBAR origins)
Fireflies Gossip
• Sign data for integrity• Even if large fraction of participants do not
forward gossip, still O(log N) completion time• But Byzantine members can launch a “reverse”
Denial-of-Service attack– By claiming to know nothing– Causes correct members to xmit all state– Root cause: lack of connections
Solution: partial membership gossip with persistent connections
Erdös & Rényi: if probability of link is high enough, graph will be connected with high probability
Chung & Lu: such a graph will have logarithmic diameter
Pseudo-Random Mesh
• Each member obtains public key certificate from CA
• Member’s id is verified by CA• Each member calculates k
positions using a secure hash function– H ( id, ring ) position
• Connect to successor and predecessor
k rings:
k chosen such that correct members form a connected graph with high probability
Working out the math…
k : out degree
N : correct + Byzantine members
n : correct members
: probability of graph being connected
But what about crashed members???
Fireflies Membership
• Members are correct, crashed, or Byzantine(Cannot necessarily tell Byzantine from correct or crashed members)
• Correct members have a view
p, q both correct q in p’s view
p correct, q crashed q not in p’s view
With probabilistic real-time bounds…
Protocol Structure
Membership Gossip
Pinging Set Reconciliation
Deliver within
Accuse and rebut
Monitor and suspect
Exchange diffs
EACH COMPONENT ISPROBABILISTIC
Here’s the problem…
• If correct members can make mistakes and falsely accuse correct members, how can we prevent Byzantine members from falsely accusing correct members???
(Answer: we don’t have to)
Membership
Each member can only accuse its 2t + 1 successors
2t + 1 rings:
t chosen such that each member has no more than t Byzantine predecessors whp
Data StructuresNote Accusation
1 2 43 553 Node identifier
1 2 43 553 Node identifier
Ring: 3
Accuser: identifier
• Both signed by creator
• Both gossiped everybody’s got the same set
t + 1 rings enabledt rings disabled
Serial numbe
r
Membership Protocol
1. On each ring, monitor the first successor that hasn’t been accused
2. Gossip accusation if successor suspect3. If you’re accused, gossip a new note (rebuttal)
with a new serial number• And disable the corresponding ring!
4. Remove from view those members who’ve been accused for longer than 2• is the gossip dissemination time
PlanetLab evaluation setup
• Configuration– t = 12 (25 monitoring rings)– Gossip rate = 1 gossip / 3.5 seconds
• Byzantine members:– 10% aggressive + 10% passive– aggressive attacks:
• accuse at any opportunity• do not forward rebuttals
– passive attacks: • never accuse• do not forward accusations
What’s still needed?
• Mandatory Information Flow– Gossip on need-to-know basis– Eliminate covert channels
• Fair Sharing (Congestion Control)– Prevent some from using up all capacity
• Secure Aggregation– Much like electronic voting– Everybody counted exactly once
• More?
Let’s look at use cases first..
• Important timely disseminations that are undesired by some and likely to be attacked– Certificate Revocation Lists– Security patches– Political messages
• Needs availability & integrity
Use cases, cont’d
• Video Streaming on the cheap– Webcasting– tele-conferencing– Entertainment
• Needs availability & mandatory information flow & fairness & fair sharing
Use cases, cont’d
• Monitoring & Auditing– PlanetLab– Cooperative Web Caching– …
• Needs secure aggregation & flow control
Use cases, cont’d
• Replication– Files– Control data
• Needs availability & integrity & information flow
Valid Accusations
• Accusation is valid iff– Accusation is correctly signed by accuser, and– Ring is enabled by the accused, and– Accuser is a monitor of the accused:
• Accuser is direct predecessor on a ring, or• holds valid accusations for everybody in between
accuser and accused
Recursive relation (without base case) complicates correctness proof…
Fireflies Group Membership
• Group membership protocol– Tolerant of Byzantine failures– Probabilistic…– Provides and uses secure gossip infrastructure
Fireflies have on/off behavior and use mimicry to dupe and devour related species…
Assumptions
• Network can lose messages
• But correct members form a connected communication graph of fair links
• Byzantine members can collude
• But cannot break crypto, and DDoS attacks can be detected and suppressed
• Bounded Pbyzantine
Informal Insight
• Gossip implements something like an atomic broadcast (whp)– Gets around problems associated with
asynchrony, such as FLP result– Atomic broadcast can implement
consensus, which is approximately what we need here…
How to pick t ?Large enough so that the probability of having more than t
Byzantine predecessors is smaller than
= 1/N
E(problem) = 1