Thoughts on Dispatcher: Enabling Active Botnet Infiltration using Automatic Protocol Reverse-Engineering

Authors: Juan Caballero, Pongsin Poosankam, Christian Kreibich, Dawn Song

Venue: CCS 2009

Summary:
The paper describes a tool that can be used to automatically reverse engineer the protocol format of an unknown protocol. The tool can be used to deduce the format and some of the semantics of the fields in both sent and received protocol messages. It can do this reverese engineering without requiring the binaries on both sides of a protocol, which makes it useful in situations where only the client is available (such as a bot binary) or only the server.

The ideas used are quite clever and simple. The main idea is to look at each byte in the buffer that constitutes a message sent or received, and then create a dependency chain for that byte. The dependency chain is the sequence of instructions and data sources used to write the byte. The dependecy chain ends when a source for an instruction is an immediate, a constant, another memory location, or unknown (such as an arithmetic op). Then by comparing the ultimate sources for contiguous destination bytes and seeing if they are derived from the same contiguous set of source bytes, they can tell the position and length of buffers that form the fields in a message. They do this hierarchically in the buffer so they can also extract the hierarchy of fields.

They can infer some semantics by looking at where these fields are used. If they are used as arguments to or if they are return values of known functions, they can immediately get the type, and some semantics such as an IP address when used in the connect function.

Furthermore, they are able to infer loops using static and dynamic analysis and then be able to infer fields that are used to indicate offsets and lengths. And by looking at repeated bytes, they are able to infer delimiters. While arrays are inferred by assuming that arrays are usually written inside loops.

To handle encrypted messages, they use the intuition behind other work, ReFormat by Z. Wang et al. The idea is that functions that have a disproportionate amount of arithmetic instructions to other types of instructions are probably cryptographic. They use this idea to infer which buffers are unencrypted and operate on them. They have used a threshold of 20 minimum instructions for a function to be considered and a ratio of 0.55, and were able to get a false positive rate of 0.002%.

For evaluation, the authors studied the MegaD command and control C&C protocol. They were able to infer the message formats used by MegaD and were able to rewrite a message for indicating SMTP sending capabilities. They have also looked analyzed other open known protocols and compared their results using Dispatcher with those from Wireshark. They say that Dispatcher outperforms Wireshark.

In terms of resiliency, the authors claim that they target fundamental properties to avoid becoming obsolete if malware adapts. However, they do confess that if malware can detect that it is being run in a virtual environment, then their technique can be circumvented.

The Likes:

• The paper is very well written and clear.
  
• The ideas are nice and clever, and they seem to work well.
• They do not need both sides of a protocol
• Can be used to infer correctness of messages when a protocol is known
• Can make writing a Wireshark dissector easier
  
• Might be used to send erroneous information to botnet masters (I don't know how useful this is. If you know that a machine is already compromised, you can should fix the situation. You cannot take over the botnet.)
  

The Dislikes:

• Can be circumvented by detecting a virtualized environment. I don't know how easy this is, but for serious malware, this might not be a problem.
• The authors claim that they target fundamental properties, but provide no evidence or proof. It is not clear that the properties they target are really fundamental. For example, they use some known functions to infer semantics. The bad guy might ship his own libraries so as to restrict the API used. The thresholds used to detect encryption and decryption are empirical, and seem easy to obfuscate. Adaptive malware might be able to change keyword values or restructure the program to make it more difficult to infer which functions are doing what. The program might be written in a difficult to analyze way where fields of the messages are always results of arithmetic computations. The program might use complex loops with jumps and gotos that make analysis difficult.
• They claim they do better than Wireshark, but it's not clear why they made that conclusion. The dissectors made different assumptions about what delimiters are, but it does not mean they are incorrect (or maybe I just did not understand what they meant).

Thought on Making Information Flow Explicit in HiStar

Authors: Nickolai Zeldovich, Silas Boyd-Wickizer, Eddie Kohler, David Mazieres

Venue: SOSP 2006

Paper: PDF

Summary:

The paper describes, in painful detail, the HiStar operating system and the implementation of two applications on it. HiStar uses distributed information flow control to provide a platform to run untrusted code on sensitive data and provide guarantees for data leakage and integrity.

Motivation:

• Hard to reason about implications of every action by untrusted
  code
• chroot can break application assumptions
• syscall interposition error-prone
• Easier to specify policy in terms of where to where information can flow
• Unlike other IFC systems, HiStar implements standard OS abstractions on top of lower level building blocks that obey IFC
• Taint tracking mechanism does not leak info itself

The HiStar DIFC model:
There are six kernel object types in HiStar. Of these, threads are the only ones that execute code. Under Distributed Information Flow Control (DIFC), each kernel object has a label. The label describes how it is tainted. An object can be tainted in different sets of categories with different degrees. The degree of taint in a particular category places a restriction on the object. In HiStar, there are 4 levels of taint.

1. Cannot be written/modified (lowest taint)
2. No restriction
3. Cannot be untainted/exported
4. Cannot be read/observed (highest taint)

Information from one object A can only flow to another object B iff B is at least as tainted as A in every category. This is the flow to relationship. A restriction imposed by a particular category can be bypassed by objects that own that category. Ownership of a category is specified as an additional degree «. Categories are only created by threads. A thread that creates a new category is the only owner of that category and can then grant ownership to another thread or to a gate (another kernel object).

A label looks like {a0, b2,c«, 1}, which means that the object is tainted in category a with degree 0, b with degree 2, c with degree «(i.e. the object owns category c), and in all other categories with degree 1 (i.e. untainted in other categories). The labels on all objects, except for threads, are immutable and have to be specified when the object is created. Objects are created by threads only.

Threads can read restricted objects by raising their label. Threads, however, can only raise their labels to the limit imposed by their clearance label. The clearance label is an additional label on threads, that also places an upper limit on the label a thread can request for a newly created object. A created object’s label must flow to the thread’s label which in turn must flow to the thread’s clearance (bypassing owned categories).

Kernel:
HiStar uses a single-level store. All objects are stored on disk and, on bootup, the entire system state is reloaded from the latest snapshot. All objects are stored in a container hierarchy with quotas.

There are six kernel object types in HiStar: segments, threads, address spaces, containers, gates, and devices.

• Segments: segments are the simplest object type in HiStar. They are basically arrays of bytes and are analogous to other OSes’ files.
• Threads: threads execute code. They can create categories and objects, and change their own label and clearance as long as the new label or clearance can flow to the current clearance ignoring owned categories.
• Containers: All objects are stored in containers that are organized in a hierarchicy rooted at the root container with quotas. Objects are hard-linked into containers, and when they are disconnected from the root, they are garbage-collected. The existence of hard links can leak information, so objects are always specified as a (Container ID, Object ID) pair. To be able to see if a hard link exists, the thread would need to be able to read the container. Quotas on objects do not change so as not to convey information.
• Address Spaces: Each thread has runs in an address space. An address space is a mapping from virtual address -> containers, segments in them and permissions. The address spaces are used to specify code to run and virtual memory. They are created by a thread and used in launching a new thread. They are used when page faults occur to find whether a thread has permission to read some segment and to find where that segment should be mapped.
• Gates: gates “provide protected control transfer, allowing a thread to jump to a pre-defined entry point in another address space with additional privileges”. Basically, they are very similar to RPCs except that the calling thread provides the resources for execution. A thread can create a gate and give it some label (respecting DIFC). Another thread can invoke the gate and request a label and clearance. The requested label and clearance are a combination of the invoking thread’s label and clearance and the gate’s label and clearance. The thread then starts executing in the new address space (i.e. executing some other code or function) using the requested label and clearance. Return values can be stored inside the thread object itself, but there is no implicit return mechanism. The invoking thread needs to create a return gate and label appropriately (the discussion on this is quite interesting).
• Devices: Can be mounted…

User-Level Libraries:
The authors implemented a Unix-like environment using the HiStar kernel completely in userspace, and so they all respect IFC. The filesystem is implemented using segments and containers. A directory is a container with a special directory segment that contains a map of names to object IDs and a mutex to synchronize directory operations. To get a consistent view of the directory entries without acquiring the mutex (which needs write permissions), a user atomically reads a generation number and busy flag before and after reading an entry.
Processes in HiStar are containers with two process-specific categories to protect its secrecy and integrity. The process has two containers, a process container and an internal container. The process container exposes objects that are the external interface to the process. The internal container has segments used to implement the file descriptors, stack, heap, and executable code. It also has the thread that executes the code and its address space.

A user has a pair of unique categories that define her read/write privileges. HiStar does not support ACLs.

HiStar also supports IPC using gates, signals, and networking using lwIP. The network device has a label {nr3, nw0, i2, 1} to taint all data read from the network.

Unix leaks information for process exit, quota adjustment, and file creation. For these operations, HiStar uses untainting gates that explicitly allow information to leak.

Applications:
The authors used ClamAV as the motivating example and split it up so that the scanner part cannot export user data.

They also implemented an untrusted user authentication mechanism that cannot leak more than one bit of the user’s password. The discussion for this application is interesting, but too complicated to explain in a summary. The main idea is to split up the login mechanism into a directory and gates and use each gate with different privileges to get the information for authentication, check the password, and get the user’s privileges.

Another application was VPN, where another lwIP stack was run in parallel to the network with a different label so that public Internet data and VPN data do not mix (at least not directly).

Performance:
HiStar runs comparable to Linux and OpenBSD, sometimes fasters, others slower. It is slower in the fork and exec microbenchmark. However, it does allow new consistency primitives such as group sync where no changes are recorded until an application is done execution.

Limitations:
Some useful features are missing from the Unix environment, while the semantics of others are modified. A crash or killed process can leave mutexes locked, and there needs to be support for cleaning those up. There are no CPU quotas and HiStar does not solve timing covert channels.

The Likes:

• The paper was thorough in the description and did not leave questions
  unanswered.
• DIFC seems very useful, and the OS seems to implement it well.
• It is nice to build a layered system basing the design on a few basic blocks that enforce security and be able to reason about the rest of the system from those.
• The OS has a load of cool and smart things whose composition was carefully designed to build useful libraries.
• It works!
• For non-malicious code, HiStar seems to be able to protect against bugs.
• It’s cool that the taint degrees work out to map into numbers that can be compared using simple integer comparisons.

The Dislikes:

• The paper was very dense, and takes quite a bit of effort to read. It would’ve been nice if the authors could have focused on the important points and insights instead of just a spewing out of information.
• The userspace environment is not completely compatible with Unix, but ah well.
• The timing channels problem seems to be a major issue
• It looks difficult to build applications using the HiStar primitives.
• What is the killer app for HiStar? What can it be used for given all its limitations? What are all these features good for?

Thoughts on The Case for OneSwarm

I should've posted this entry a while ago. This is a talk that Tom Anderson gave at MSR in Mountain View on 9/29/2009.

Summary:

Cloud computing has many issues:

• Data can be locked in (Facebook, ...)
• Need to trust providers with data
• Cloud is a natural monopoly:
• The business model is to attract developers then the users. So the users have no choice.
• Reliability and the ability to handle flash crowds is a high barrier to entry.
• There is no plausible avenue for an open source Cloud. Somebody needs to pay for the infrastructure, so somebody needs to be making money off of it to be sustainable, and the way to do this is through data lock-in and/or data mining...
• Since all data is stored somewhere centrally, it's easier to censor it

Can we use a P2P system instead where users can donate some of their resources to be able to use other people's resources? At first glance, it looks like this might be alright. P2P systems

• have high availability -> But in fact this is not true because seeds disappear
• have no bottlenecks -> But their could be network inefficiencies or ISP throttling
• have no centralized monitoring -> But it's actually easy to track users

The alternative is OneSwarm. OneSwarm has no centralized trust. The idea is to use privacy-preserving sharing where you build a web of trust through friends. Then when a user makes a query, the query goes through friends, then friends of friends, and so on... and the data travels on the reverse path back to the user. Nodes are not supposed to reveal information about where a query they pass along came from, so the receiver of a query cannot relate a user to a query.

The system, is by Tom's admission, vulnerable to collusion, but that's okay. The system is not designed for complete security but rather as a compromise between security, flexibility, and performance. Another assumption is that there is no partitioning in the network and that eventually everyone will be connected, and everyone will have access to everything (which will probably end up happening anyway).

The network of friends is organized as a DHT. To maintain availability and solve the issues with traditional P2P systems, they invented a new type of DHT that is highly scalable and reliable. groups of nodes are organized into tight groups that are replicated via Paxos. Each group then appears as a single node in the DHT.

As a final note, I had a chat with Tom after the talk. He has a pessimistic outlook on what will eventually happen. Mainly, that cloud providers will always end up creating a monopoly using data lock-in and be able to mine the data stored on the cloud, because they will manipulate the API so they can do so.

Likes:

I like OneSwarm. In particular, I liked the DHT implementation. They had some really cute ideas there. It is not a security protocol, it is a "casual" privacy protocol that presents a minimum barrier to snooping. However, if someone wants to get to my data and to know what requests I am making, they would be able to do so.

The availability of a P2P system that is presented as cloud storage on top which users can implement "free" applications, such as an open-source Facebook.

Dislikes:

• I'm not sure I agree with the pessimistic view. Google has started a movement called the data liberation front to basically free users from data lock-in, and I expect this movement to become more important.
• There are more reasons than just convenience for applications to run in a data center. If we move from DCs to P2P systems for storage, we will get much higher lookup costs, and it is not clear systems will scale.
• There are no guarantees of connectivity. The social network might get partitioned, limiting the available data.

Thoughts on A System for Authenticated Policy-Compliant Routing (Platypus)

Authors: Barath Raghavan and Alex C. Snoeren

Venue: SIGCOMM 2004

Paper: http://www.cs.ucsd.edu/~snoeren/papers/platypus-sigcomm04.pdf

Summary:

Providers want to forward packets only for entities they can bill. Regular source-routing makes this difficult since packets can arrive from anywhere and it is not clear whom to bill. The authors propose a network protocol that allows an ISP to give capabilities to whomever it wants to bill. The capabilities themselves can then be further delegated. Their work assumes that the knowledge of alternate routes or waypoints is done through some out-of-band means and the issue is not addressed.

Applications:

1. Efficient overlays and on-demand routing where end-hosts can decide how to route packets.

2. Preferential egress points where different peers get routed differently even if they all peer at the same router.

3. Preferential ingress points where multi-homed ASes can select what which ISP traffic comes in through.

4. Virtual multi-homing where a stub AS can get multi-homing at the provider (i.e. a hop or more away) even though it only has one connection.

The Platypus header is below the IP header so it is compatible with IP. A packet has a list of capabilities where each capability specifies the resource principal to bill, and the waypoint to go to. Packets are forwarded between waypoints by replacing the IP address in the IP-level packet.

Capabilities are authenticated using a “double MAC” trick. The key server and the router share a single key k. Secret capability key s = MAC(k, capability) is given to customer. Customer then generates b = MAC(s, payload) and puts b in the packet along with the capability. The router can then authenticate the capability.

The capability key s is temporary and expires after some time. The expiration mechanism used requires loose synchronization, but allows for only one value. When a key expires, the customers need to get a new key. Customers are given a master key they can use. They encrypt the op code for requests and put it in the DNS lookup name. So if a key is requested from a server a.b.edu, the request is sent to E(master key, request code).a.b.edu. The response contains the encoded new key. Requests are thus cacheable.

They also have revocation lists that are pulled from key servers into the routers.

The describe two ways to do delegation on prefixes. One by XORing, and the other by recursive MACing. The first is susceptible to collusion, the second might be too slow (this is what we do in ICING).

They evaluate their system using kernel modules, and it is a little slower than IP forwarding. I did not inderstand their discussion on clustered waypoints and their proposed solution traffic engineering. So I won’t say anything about it. I will say that the way they choose their nonces is not very nice.

They punt issues like accounting saying it is possible to do some distributed per-packet counting and rate-limiting(unknown). Their scalability section is a little hand-wavy since they have no idea what it would really take to build it in hardware. Replay attacks are also hand-waved saying they will use bloom filters (but they don’t mention costs or how they would do that in hardware, what do they hash…)

The Good:

· This is good seminal work on how to verify policy compliant packets.

· Used some cute smart tricks here and there (expiration, double MAC, delegation, key lookups)

· System seems to perform well enough in software

· They present the system and crypto nicely (who has what keys and how they derive them for example)

The Bad:

· They concentrate only on recognizing whom to bill. In reality, policies are probably more complex for security reasons. For example, don’t let AT&T send packets through Sprint even if another customer is paying for it. Or only use trusted ISPs.

· Much has been hand-waved. The implementation doesn’t explicitly specify (or maybe I missed it) what has been implemented.

· Revocation lists and polling from key-servers might be a problem.

· Once a capability is in a packet, it is compliant. There is no check whether or not the packet actually follows the path.

Thoughts on HadoopDB: An Architectural Hybrid of MapReduce and DBMS Technologies for Analytical Workloads

Authors: Azza Abouzeid, Kamil Bajda-Pawlikowski, Daniel Abadi, Avi Silberschatz (Yale University), Alexander Rasin (Brown University)

Venue: VLDB 2009

Paper: http://db.cs.yale.edu/hadoopdb/hadoopdb.pdf

Summary:

The paper starts by saying that DBs are becoming increasingly important and data is growing at a huge rate. Unfortunately, the usual parallel DBMSs do not scale well because they cannot handle failures very well. On the other hand, MapReduce handles failures and stragglers (heterogeneous systems) well and scales well, but does not perform as well as DBMSs.

The paper describes a system that has roughly three layers. The top layer is a SQL layer which accepts SQL queries from users. The SQL queries are translated into MapReduce jobs. The MapReduce jobs are executed in the middle layer, Hadoop. The job is divided into several tasks. Each task itself is a smaller SQL query that is executed on a single node. Each node independently runs the third layer, a single-node DBMS (PostgreSQL in this case). It uses Hadoop for communication, Hive for the translation, and PostgreSQL for the DB.

HadoopDB consists of the following main components:

1. The database connector: provides an interface to independent databases so that worker nodes (TaskTrackers) can get results of SQL queries in MapReduce format (i.e. key-value pairs).

2. Catalog: maintains metainformation about the third-layer DBs---connections parameters (e.g. DB location, driver class, credentials) and metadata (data sets, partitions, replica locations, …)

3. Data Loader: Partitions data and stores it in the local DBs.

4. SQL to MapReduce to SQL (SMS) Planner: Extends Hive which translates SQL queries to MapReduce (Hadoop) jobs by optimizing the output to push as much of the work onto the local DB instances in the worker nodes since the single node DBMSs can do much better than MapReduce.

In the eval section, they compare Vertica (a column store DBMS), DBMS-X, HadoopDB, and Hadoop. They tested the systems without failures and with failures. I didn’t quite understand all the differences in workloads, but I did notice that Vertica was much faster than any of the other systems. The authors claim that HadoopDB performs almost as well as the other systems they tested and almost always better than Hadoop. They say that the performance differences are mainly because of the lowest PostgreSQL layer. This seems plausible, but it would have been a much more compelling eval if they tested comparable systems (example use compression on PostgreSQL and/or a column store).

When faults and slow nodes were introduced, Vertica performed much worse: 50% performance hit with failures and 170% with slow nodes. HadoopDB on the other hand was only 30% slower. They then say that this slow-down is a minor bug that can be easily fixed so that their performance matches Hadoop’s (which is better than HadoopDB with slowdown).

The Good:

The paper is interesting and reads well. The system itself is also very smart, and it’s a wonder no one has built it before. I would really like to see what ends up happening with it and how it ends up performing in the real world.

The Bad:

It would have been much more insightful if the used the same DB bottom layer in the eval, or at least eval different ones. It would have been compelling to see how Hadoop can parallelize DBMS. It would have been nice too if they could have put Vertica there to show how HadoopDB can fix the issues with failures. Also, I didn’t think the queries the used were that complex, but I don’t know whether this is of any importance.

Thoughts on Fabric: A Platform for Secure Distributed Computation and Storage

Authors: Jed Liu, Michael D. George, K. Vikram, Xin Qi, Lucas Waye, Andrew C. Myers (Cornell University)

Venue: SOSP 2009

Summary:

The paper describes Fabric: a language that is mainly an extension of Jif into the distributed computating world. The paper start with motivation citing patient record sharing as an example of where it is necessary to share data between two domains such that each domain's security policies are enforced.

In Farbic, there are three types of nodes:

1. Stores: store objects persistently
2. Workers: perform computation
3. Dissemination: CDNs for stored objects

Naming: Fabric uses an object oriented DB model. Fabric objects store information, are mutable, and have an object id or oid. The oid is mainly the concatenation of the location where the object is stored (as a DNS name) and a 64bit number. The oid is permanent. If the object moves, a pointer in the old location will point to the new location.

Certificates are the roots of trust and are used to establish SSL connections.

Labels: Objects has labels decribing its integrity and secrecy requirements.

Classes: Class objects creat an unforgeable binding between each object and the correct code for implementing that object.

Versions: Fabric objects are mutable. Each has a version number that is incremented when a transaction updates the object. The version number is used to check whether a transaction is operating on a stale or the up-to-date copy of the data.

Threat Model: Timing and termination channels are ignored. It is assumed that the network will deliver messages eventually. It is not clear if this assumption is a strong one. Do they need the network to never fail or else everything crumbles down? Or do the only assume that the network works mostly as today (it will be out for a few hours maybe)?

Another assumption that is used in the rest of the paper is that a program knows a priori the nodes it will use for computation and dissemination, or that they will be input into Fabric somehow. Similarly for storage nodes.

Storage Nodes: Store objects. Related objects are grouped together and returned when requested for efficiency. They track delegation or as they call it "trust relationships".

Worker Nodes: Execute Fabric programs.

• Trusted programs can incorporate other intermediate language-programs that don't have security constructs.
• Objects modified only inside transactions
• worker nodes can either request remote data (data-shipping)
• or can do remote invocations (function shipping)
• Remote invocations don't need the object to already exist at the remote worker node
• The entire method call is executed in a subtransaction whose effects are only visible when the top-level transaction commits. The whole transaction tree is then committed.
• The caller-side of the remote invocation checks its security statically. The callee-checks it on run-time since it doesn't know who is going to invoke it. i.e. the langauage has to explicitly check trust: e.g. if a trusts b: then a.invoke(...)

Dissemination Nodes:

• Dissemination layer is not prescribed, and currently uses FreePastry
• Object on dissemination nodes encrypted
• Key stored on object's store
• Store can verify security and dole out the key when safe. **They say workers need not fetch keys often, but not clear why...p5
• Stores and dissemination nodes track readers.
• When object updated, updates are pushed to tracked worker nodes.
• Dissemination nodes can learn info from request pattern, so they are labeled with some label representing "the maximum confidentiality of information that may leak on this channel". Workers only do fetch requests if this max is not violated. **I do not understand this whole paragraph. How do they quantify the amount of information leaked on this covert channel and turn it into a label?

The Fabric Language: They extend JIF to support

1. Transactions and consistency
2. Remote method calls

The interesting thing is they do both of these without leaking information and so that remote calls are authorized. I think this should probably say "where information leakage is explicit" because there will always be some undesired information leakage.

Then they go on to say: "Fabric adds a new trust ordering on information flow labels". This I found confusing and intriguing at the same time. Eventually, I will be disappointed...

In Fabric, everything is in terms of principals.

• There are no partial privileges or dynamic labels for prinicpals. A principal always executes with its full privileges.
• They add an acts-for relationship which is basically just delagation of all privileges, and they call this "trust". If a acts for b then b trusts a or b has delegated all its privileges to a.
• There's a top principal (I'll call T, but they use a symbol I'm too lazy to lookup) that acts for all other principals.
• There's a bottom principal (B) that all prinicpals act for.
• Fabric nodes are also principals, and so also follow the acts-for model.
• Principals implement their own authentication. The description of how trust is bootstrapped in the system is too confusing. They must have cut too much here. My imagination fails me. **They need to describe bootstrapping more. How does a user delegate trust to a node? How is the authentication turned into a principal?
• Information flow security policies are expressed in terms of principals. They claim this integrates access control and information flow, though I don't know what that means. Information flow security policies are access control policies!
• Labels in Farbic look like: {alice -> bob} meaning alice, the owner of this security policy allows bob to read from it. {alice <- bob} means alice allows bob to write to it. Since the owner is explicit (alice), only code running with alice's authority (worker node acts for alice) can declassify or endorse data with alice's labels.
• {alice->bob} can flow to {charlie->dora} only if charlie acts for alice and dora acts for bob or dora acts for alice.

Trust Ordering: As well as being ordered by information flow using the flow to relationship, labels can be ordered by trust (or the acts-for relationship). The higher a label's trust requirement, the more a node needs to be trusted. For example, {a->b} actsfor {c->d} iff a actsfor c and b actsfor (d or c). That is one label can be used instead of another in that case. This can be used to establish trust in nodes: if an object x with label L_x is to be stored on a node n. Formally, L_n={T->n; T<-n} actsfor L_x for this to work.

e.g. if L_x is {p->q} then either n actsfor p or n actsfor q. In either case, n must be trusted with the contents of x.

They mention that the reason this trust ordering is needed is because "Fabric classes may be parametrized with respect to different labels or principals, and so instances of the same class may have different labels. This allows implementation of reusable classes..." However, it is not clear to me that anything other than good ol' delegation is necessary. Of course you have to delegate to a node if that node is to enforce your labels! That is why in D-Star exporters own categories they in labels they enforce. Hence, the disappointment I alluded to earlier.

As in Jif, implicit flows are tracked with the program counter's label.

For remote invocations, they say that the callee checks at run-time that the caller has the authority to invoke something with another principal's endorsements, but they do not mention how this check is done. Are there certs as in D-Star? **Need to mention how delegation is proved across nodes.

Transactions: Fabric provides transactions that can be distributed. When distributed, they are turned into a transaction tree. The whole tree is committed when the top level worker node starts the commit. It does so by recursively calling the commits of the whole tree (note recursively not iteratively). Most of the things they describe about transactions are how you would expect them to work.

A failure of a node to commit a transaction correctly is as serious as if that node violated trust by not updating the object's fields correctly.

If a transaction coordinator midway through a transaction, so changes might be committed while others not. This can break consistency but only for objects whose integrity the failing node was entrusted with.

Writer Maps: Used to allow workers to check whether an object has been updated without revealing information to workers who shouldn't learn about the updates. The writer map a log of writes. Everytime an object is updated, an entry hash(oid, xid, key)->E(key, w) is stored where w is the writer. It allows nodes that know the key to get the latest update and know what the writer for that update is. But it is not clear how the nodes know what the last transaction id is. It is probably stored in the clear. **Nodes also still know that an object has been updated since the writer map changes even if the are not allowed to know what the exact updates are.

Implementation and Eval: The implementation was 33k lines of code on top of Jif and other libs. They use BDB for the store and FreePastry for dissemination.

A few things are left unimplemented, most notably the distributed deadlock detection and timeouts for divergent computations.

For the Eval they implemented a course management system and tested on 55 students. Complex queries were replaced with object oriented code and this simplified the implementation from 3100 loc to 740. However, they only migrated part of the full system. They also only ported it to FabIL, the intermediate language used by Fabric that has no security properties. Only consistency. Performance shows that FabIL is faster than all other implementations with consistency. Transaction management and consistency is about 50% of performance. **The metric doesn't really show much since only parts of the implementation were done. It is not clear why FabIL is faster. Maybe I didn't understand what FabIL does. Is the implementation on a set of nodes? How does BDB run faster that Oracle? Still it says nothing about security and performance with full Fabric used.

They also implemented a calendar. They didn't get performance numbers. Is it because the calendar didn't fully work?

They used the OO7 benchmark on Fabric nodes and found that Fabric is 10 times slower than non-persistent Java. They say this is acceptable, but that is not clear.

In summary, the paper is a thorough design of a language for distributed computation and storage when DIFC is involved. But like almost every other paper on the subject has left me with more questions than answers.

The Likes:

• The paper is thorough in the design description
• The paper is probably the first to combine DIFC with transactions to provide consistency
• They combine language constructs with run-time checks. It's cool that trust requirements are in the language.
• They do so much stuff!

The Disklikes:

• Ah so many questions! See the text (**). I'm too lazy to put them here.
• Why the trust ordering on labels? Why is delegation not enough? I am not convinced that it is necessary.
• They have said almost nothing about how runtime checks are implemented. Do they use certs as in DStar to prove delegation?
• That bootstrapping paragraph was completely useless. It left me more confused than informed. Probably because they did not explain how runtime enforcement is done and how authentication is turned into labels or proofs of labels.
• The evaluation leaves so much to be desired. There is no full implementation of any system that uses Fabric and no real performance evaluation. In all, that was a disappointing eval section.
• What is the deal with the dissemination nodes. Yup they are cool, but where is how they help? Are they basically CDNs? How are they used? Why not just replication?
• I don't like the fact that principals always use their full privileges to access everything. What if a principal wants to operate in an untrusted environment or where it doesn't want a node to be responsible for the integrity of so many objects.
• The paper does not say much about the "control plane": Where do the names of nodes come from? How does a user select nodes? How do programs select/find nodes? How does a user's credentials translate across multiple nodes? I guess all this is orgthogonal but it would've helped me understand Fabric's usage model which was quite obscure (as in many other DIFC papers).
• I think my main issue is why is the system not layered? Why aren't transactions built on top of a minimal base whose properties are easy to reason about? Why are dissemination nodes a primary object instead of being implemented on top of worker and storage nodes, i.e. on top of Fabric? Is it convenience? What happens when things need to change? What about the modularity argument?

Thoughts on Provable Data Possession at Untrusted Stores

Authors: Giuseppe Ateniese† Randal Burns† Reza Curtmola† Joseph Herring†
Lea Kissner ‡ Zachary Peterson† Dawn Song §

†Department of Computer Science, Johns Hopkins
University, Baltimore, MD – {ateniese, randal,
crix}@cs.jhu.edu, jrh@jhu.edu, zachary@cs.jhu.edu
‡Google, Inc. – leak@cs.cmu.edu
§University of California Berkeley/Carnegie Mellon Univer-
sity – dawnsong@cs.berkeley.edu

Venue: CCS’07, October 29–November 2, 2007, Alexandria, Virginia, USA

Paper: http://www.cs.berkeley.edu/~dawnsong/papers/p598-ateniese

Summary:
The authors describe a way to probabilistically determine whether a file stored at a SSP exists and has not been modified. I don't quite understand the crypto, but the results are:

• Fast checks. They are I/O bound, not CPU bound
• The size of the file stored at the server increases
• The client needs to store a constant small amount of data per file
• They use sampling to reduce the load on the server.
• Uploads are slow because preprocessing is expensive on the order of 160kB/s

This work is almost identical work by Ari Juels(RSA Labs) and Burton Kaliski(EMC Corp) PORs: Proofs-of-Retrievability for Large Files. However, in PORs, a client has to decide the number of checks a priori because the system works by encoding a constant number of sentinel blocks for checks in the file that are queried and hence consumed during challenges.

The main difference between PORs and PDPs is: PORs prove to a client that the entire file can be retrieved unmodified. PDPs do not make this claim. They only claim the file blocks tested are uncorrupted.