Andreas Haeberlen †‡ Petr Kouznetsov † Peter Druschel †
† Max Planck Institute for Software Systems ‡ Rice University
Systems can be protected with Byzantine fault tolerance (BFT) techniques, which mask a bounded number of Byzantine faults, e.g. using state machine replication . BFT is a very powerful technique, but it has its costs. In a practical system that needs to tolerate up to f concurrent Byzantine faults, BFT cannot be implemented with less than 3f+1 replicas . Moreover, BFT scales poorly to large replica groups; as more servers are added, the throughput of the system may actually decrease .
In this paper, we explore an alternative approach that aims at detecting rather than masking faulty behavior. With this approach, the system does not make any attempt to hide the symptoms of Byzantine faults. Rather, each node is equipped with a detector that monitors other nodes for signs of faulty behavior. If the detector determines that some node has become faulty, it notifies the application software, which can then take appropriate action. For example, nodes can cease to communicate with the faulty node; once all correct nodes have followed suit, the faulty node is isolated and the fault is contained.
Fault detection is weaker than masking. For instance, detection is insufficient for dealing with faults that have serious and irreversible effects, such as deletion of all copies of an important document. However, detection may offer an efficient and scalable alternative to BFT for faults that have limited or recoverable effects, including freeloading, censorship, and denial-of-service.
We are interested in fault detectors that provide accountability . With such detectors, each action is undeniably associated with the identity of the node that has performed the action, allowing the system to gather irrefutable evidence of faulty behavior.
The fault detection systems we consider should guarantee at least two properties. The system should be complete: whenever a correct node observes the effects of faulty behavior, the system eventually generates evidence against at least one faulty node. Also, the system should be accurate: it never generates valid evidence against a correct node.
Adding accountability to a distributed system has several important advantages, regardless of whether the systems uses BFT or not: first, any faulty behavior by a node is guaranteed to be detected. Second, the evidence produced by the detector can be used to convince third parties that a fault has occurred. Third, the evidence enables the system to resolve he-said-she-said situations in which two nodes accuse each other of being faulty. Lastly, the presence of accountability alone deters certain types of attacks on a system, because it identifies and exposes faulty nodes.
Our goals in this paper are threefold: first, we examine the trade-offs between fault detection and traditional BFT. Second, we give a precise definition of the class of Byzantine faults that can be detected with our approach. Finally, we briefly sketch a practical system that implements a Byzantine fault detector.
Techniques that mask Byzantine faults are perhaps easier to use than fault detection systems, since they provide the application designer with the abstraction of a system in which failures simply do not occur. So what reasons are there to opt for fault detection?
Detection requires less replication: If a practical system can suffer up to f concurrent faults, BFT cannot be implemented with less than 3f+1 nodes . As we will show, detection can be accomplished with only f+11. The complementary view of this point is that BFT requires the fraction of faulty nodes in the system to remain below 33% at all times, while a correct node can reliably detect faults irrespective of the fraction of faulty nodes.
Detection systems need only be provisioned for the average load: In a BFT system, all replicas must process each request promptly, since the client cannot make progress before most of them have responded. In a detection-based system, however, a single replica can process each request and respond immediately; the other replicas can later check the response during a period of light load. Hence, a BFT system must be provisioned such that each machine can handle the peak load, while in a detection system, each machine must merely be able to handle the average load.
Detection is cheaper: Detection avoids the consensus required by BFT, and it enables extensive aggregation of messages, state and processing associated with detection. Also, there is no need for strong consistency among the replicas, which makes it much easier to handle changes to the replica group.
Detection systems do not only require fewer resources than BFT, they also have some functional advantages that benefit distributed systems whether or not they use BFT:
Detection enables a timely response to faults: In a system that does not use BFT, once correct nodes obtain evidence of a fault, they can stop communicating with the faulty node and thus isolate it. Also, correct nodes can initiate recovery, e.g. by creating additional replicas of any objects affected by the fault, or by alerting a human operator who can repair the faulty node. Timely repair can also help BFT-based systems to stay within their bound for the number of concurrent faults.
Detection provides a deterrent: The mere presence of a detection system can reduce the likelihood of certain faults. For example, it can discourage freeloading and censorship in peer-to-peer systems by creating a disincentive to cheating, since a faulty node risks isolation and expulsion from the system. Furthermore, if the system maintains a binding from node identifiers to real-world principals, then the owner of a faulty node can be exposed and held responsible. Reducing the frequency of certain faults also benefits BFT-based system, allowing it to more easily maintain its error bound.
Next, we sketch general application areas for fault detection systems, along with some specific examples.
Systems with recoverable state: Network file systems and distributed information systems typically perform periodic backup snapshots to ensure data durability. It is usually acceptable for these systems to revert to the latest snapshot in case of a serious malfunction or attack. However, faults must be discovered quickly in order to prevent corrupt data from spreading to the backups. By adding a detector, these systems can bound the time to detection for a very general class of faults.
Redundant systems: Decentralized systems and systems based on BFT mitigate or mask the effect of a limited number of faults through redundancy. However, if faulty nodes are not discovered and removed quickly, they can accumulate over time and eventually lead to a system failure. Using a detector, faulty nodes can be identified and isolated before they can cause any serious harm.
Systems that span multiple administrative domains: Such systems can benefit from accountability to keep the players honest and to apportion blame, e.g. in federated databases and hosted Web services, or to discourage freeloading and censorship in peer-to-peer systems. The Internet's inter-domain routing system is another example. In the case of a malfunction, detectors could not only identify which party is at fault; they would also produce evidence that would allow other parties to prove to their customers that they are not to blame.
Before defining formally the classes of Byzantine faults that can be detected using observable messages, we discuss a simple example protocol that has only two methods: A put method, which is used to store an object on a node, and a get method, which is used to retrieve it. Figure 1(a) shows a simple example of a message exchange in which node B receives an object from node A and later delivers it to another node C.
Now assume that node B is faulty and wants to prevent node C from obtaining the object. There are two ways for B to achieve this. One is to break the protocol and deny C's request, as shown in Figure 1(b); we call this behavior detectably faulty. The other is to pretend that it has not received the request message, as shown in Figure 1(c); we call this behavior detectably ignorant.
In both cases, the fault is detectable because it affects the message exchange observed by the correct nodes. However, a node might become faulty but continue to follow the protocol exactly as if it were correct. Such a fault cannot be detected with our approach. Similarly, if a fault is completely internal to one node or affects only messages sent to other faulty nodes, it is not observable by any correct node and therefore cannot be detected.
In the rest of this section, we provide a formal definition of detectable faultiness and ignorance, and we introduce a set of guarantees that can be given by a detection system.
We consider a set Π of nodes. Every node i is modeled as a state machine Ai and a detector module Bi (Figure 2). Informally, we say that a node i is correct if it respects the specifications of both Ai and Bi. Otherwise, the node is faulty.
Nodes communicate with each other through message passing. We assume that messages are uniquely identified. For a message m, let sender(m) and receiver(m) denote the sender and the receiver of m, respectively. For the moment, we do not put any restrictions on local processing time and communication delays. However, we assume that, after some number of retransmissions, a message sent from a correct node to a correct node is eventually received.
An event is either sendi(m)∈ Oi, where i=sender(m), or receivej(m)∈ Ij, where j=receiver(m), or an application-specific input or output.
An execution E is a sequence of events such that in E, each m is sent and received at most once, and each receivei(m) is preceded by the corresponding sendj(m). We distinguish events associated with the state machine Ai and events associated with the detector module Bi. E|Ai denotes the subsequence of E that consists of all events associated with Ai, and E|Bi denotes the subsequence of E that consists of all events associated with Bi. We say that a node i is correct in E if (1) E|Ai (respectively, E|Bi) conforms to Ai (respectively, Bi), i.e., if the sequence of outputs produced in E|Ai (E|Bi) is legal, given Ai (Bi) and the sequence of inputs in E|Ai (E|Bi) and (2) if E is infinite, then both E|Ai and E|Bi are also infinite. Otherwise we say that i is faulty in E.
We define a history of a node i as a sequence of events of Ai. A history h of a node i is valid if it conforms to Ai, i.e. if, given the sequence of incoming messages and application-specific inputs in h, Ai could have produced the sequence of outgoing messages and application-specific outputs in h. A pair (h1,h2) of histories of i is consistent if h1 is a prefix of h2, or vice versa. If i is a correct node, one trivial example of a valid history is E|Ai.
Let M(E) denote the set of messages received by the nodes in an execution E. We assume that there exists a history map φ that associates every message m∈ M(E) with a history of sender(m). For a correct node, φ(m) is the prefix of the local execution E|sender(m) up to and including send(m). Thus, for any message m sent by a correct node, φ(m) is valid, and for every pair of messages m and m' sent by a correct node, φ(m) and φ(m') are consistent.
We say that a message m is observable in E if there exists a correct node i and a sequence of messages m1, ..., mk such that
In other words, m is observable if it causally precedes at least one event on a correct node.
We say that a node i is detectably faulty with respect to a message m sent by i in an execution E if m is observable in E and satisfies one of the following properties:
We say that a node i is detectably ignorant in E if i is not detectably faulty in E and there exists a message m sent to i by a correct node, such that, for all observable messages m' sent by i, receivei(m) does not appear in φ(m').
When the detector Bi on a correct node i has seen evidence of faulty behavior on another node j, it sends a failure indication to its local application process. We define three different types of indications: trustedj, suspectedj and exposedj. Intuitively, if the detector Bi outputs suspectedj, there is evidence that j is ignoring certain inputs, e.g. by refusing to accept a service request from a correct node. If it outputs exposedj, there exists a proof that j is faulty, i.e. that it has deviated from the specification of its state machine Aj. Finally, Bi outputs trustedj while none of the other conditions hold.
We can use a definition similar to that of [5,11] to describe these properties. Thus, the detection system guarantees that the following properties hold in every execution:
Note that the detector need not guarantee that a correct node is always trusted by another correct node; it can jump from trusted to suspected and back, e.g. due to long message delays. Further, if a set of colluding faulty nodes includes at least one detectably faulty node, then at least one of them will eventually be exposed or suspected forever; we chose this weaker guarantee to facilitate an efficient implementation of φ. Nevertheless, if there are only finitely many faulty nodes in the system, correct nodes can be affected by their behavior only finitely long.
For PeerReview, we assume that the system can be modeled as described in Section 2.2, with two additional assumptions: First, the protocol is deterministic, i.e. it produces the same outputs given the same sequence of inputs. This is a fairly common assumption in state machine replication [12,4]. Second, nodes have strong identities and hold a cryptographic keypair that can be used to sign messages. This can be accomplished, for instance, by giving each node an identity certificate, signed by a certification authority, that ties its public key to its node identifier.
We also make the common assumption that the attacker does not have the ability to break cryptographic signatures. Other than that, the Byzantine nodes may behave arbitrarily and collude with each other.
Each node is required to keep a log of all the inputs and outputs of its local state machine Ai. The log is organized as a hash chain, similar to a secure history , such that the top-level hash covers the contents of the entire log. Furthermore, each node must frequently commit to the contents of its log by publishing an authenticator, i.e. a signed copy of its top-level hash value. This makes the log tamper-evident and ensures that nodes cannot revise their history .
Nodes must sign all messages they send and acknowledge all messages they receive. If a message is not acknowledged after several retries, it is broadcast to the other nodes, who then challenge the node to accept the message. This ensures that a node is suspected by all correct nodes if it refuses to accept a message.
Each message or acknowledgment m contains an authenticator, as well as a short proof that send(m) or receive(m) was the top-level entry of the corresponding log. The recipient extracts the authenticators and, once in a while, forwards them to the other nodes. Thus, all interested nodes eventually learn of all authenticators that have been signed by a correct node.
Each node i is periodically audited by other nodes. During an audit, the auditor j first asks i for a signed log segment that covers all entries since the last audit. j then validates the log against the most current authenticator it has obtained for i. If i refuses to comply, j begins to suspect i.
Next, j performs a consistency check to see if the log matches all the recent authenticators it has obtained for i. If two or more authenticators do not match, then i has either forked its log or is keeping multiple copies. The mismatched authenticators are then made available to other nodes as evidence, who can thus mark i as exposed.
In a third step, j extracts all authenticators from the log segment and forwards them to other nodes. This ensures that, even if i is faulty, other nodes will eventually be aware of all relevant authenticators.
Finally, j performs a conformance check. It instantiates a local copy of the state machine Ai and initializes it with a recent checkpoint from the log. Then it replays all the inputs from the log and checks whether the corresponding outputs match the ones in the log. Thus, j can check protocol conformance without an explicit protocol specification. If it detects a divergence, it has obtained a signed confession and can thus expose i.
In a naive implementation of the algorithm, every node audits every other node, requiring O(N2) messages and computation. Though the messages are small and can be heavily aggregated, this does not scale to large systems. In practice, however, if at most f nodes can be faulty at any given time, it suffices if f+1 different nodes audit a given node, which reduces the complexity to O(f2). The details of this optimization and a full evaluation will be provided in a subsequent full paper.
If a node j detects a fault on a node i, it obtains one of two types of evidence. If i is detectably faulty, j obtains either a) an authenticator and a log, both of which are signed but do not match, or b) a signed log segment that fails the conformance check. Both constitute a signed confession. If i is detectably ignorant, j obtains a challenge (e.g. a request for a certain log segment) that i cannot answer, except by providing a signed confession.
Both types of evidence can be distributed to the other nodes, who can verify them independently, either by repeating the checks performed by j (in case of a signed confession) or by contacting i and checking its response (in case of a challenge). PeerReview ensures that such a check will always fail for a correct node, since it never produces a signed confession and can respond to any challenge.
The output of the PeerReview failure detector on a given node is reliable if, and only if, the node has a valid copy of the state machine to be run by all the nodes in the system. A node can ensure this, for instance, by obtaining a signed binary program from a trusted authority.
To bound the space required for logs, nodes may be allowed to discard old log entries, e.g. after a month. In this case, older evidence can no longer be verified and must be discarded as well, which eventually allows faulty nodes to return to the system. This is acceptable as long as the system has ample time to respond to the failure and initiate repair. If repair is not an option, e.g. in a large decentralized system, the log must be kept long enough to create a serious disincentive for attackers. Alternatively, the remaining nodes could use Byzantine consensus to permanently revoke faulty nodes' certificates prior to truncating the log.
Our concept of a detection system is based on the failure detectors by Chandra and Toueg . These were defined for crash failures, but Malkhi and Reiter  later extended them to special classes of Byzantine failures. In a more general manner, Doudou et al.  have introduced muteness failure detectors dealing with nodes that prematurely stop sending algorithm messages. Kihlstrom et al.  have introduced several classes of failure detectors that expose detectable Byzantine failures. However, they consider classes of algorithms in which all messages are broadcast, and in which processes know when to expect messages from other processes. PeerReview does not require these assumptions.
State machine replication [12,16] is a classical technique for masking a limited number of Byzantine faults. Today's state-of-the-art BFT techniques, e.g. , are based on this idea. Although these techniques perfectly protect the system from Byzantine failures, they are usually not intended to detect the faulty nodes, and they are inherently expensive and not scalable. The BAR model  extends the BFT approach to tolerate selfish behavior of rational nodes while providing a mechanism for detecting certain application-specific misbehavior. Our approach is more general and, unlike BAR, it does not inherit the algorithmic complexity of BFT.
Alvisi et al.  introduced a technique that monitors quorum systems and raises an alarm if the failure assumptions are about to be violated. This technique is probabilistic and, unlike PeerReview, cannot identify which nodes are faulty.
Intrusion detection systems  can detect certain types of protocol violations; however, unlike PeerReview, the heuristics used in IDS tend to produce either false positives, false negatives, or both. Reputation systems such as EigenTrust  can be used against Byzantine failures, but, unlike PeerReview, they cannot prevent a coalition of malicious nodes from denouncing a correct node. Finally, trusted computing platforms like TCG/Palladium can detect failures that involve software modifications, but force users to exclusively run certified software. PeerReview merely checks protocol conformance but otherwise allows diverse implementations.
Yumerefendi and Chase  proposed accountability as a first-class design principle for dependable distributed systems, but mentioned that general, application-independent techniques were still elusive. We believe that detection systems, and PeerReview in particular, are a major step towards this goal.
Detection promises to reduce the cost of robustness to Byzantine faults, and to increase dependability of systems in which BFT is infeasible or prohibitively expensive. For example, detection offers a relatively efficient defense for freeloading and censorship attacks in large-scale distributed systems. It also enables accountability in systems that span multiple administrative domains, such as federated databases, hosted web services, peer-to-peer systems and Internet inter-domain routing.
We believe that further research in detection systems will yield a variety of new detectors with different tradeoffs. For example, more powerful detectors could be constructed by adding more sensors, such as attestation, and hybrids between detection and BFT could allow more fine-grained tradeoffs between protection and overhead.