August 24, 2016

Archives for February 2006


Obligatory Summers Post

According to Section 4.3(c)(iii)(g) of the Code of Academic Blogging, I am required, on pain of banishment from the faculty club, to post about the departure of Lawrence Summers as Harvard president. Much e-ink has been spilled on this topic, and I for one feel no wiser for it. With some trepidation, let me offer a few thoughts.

I am the first to admit that I don’t know much about how to run Harvard, or about what kind of job Summmers did. I read the same press articles as everybody else, but in my experience outsiders commenting on university matters hugely overweight things they read in the newspapers, and underweight less flashy details of management. For example, I as a Princeton professor don’t much care what Princeton’s president thinks about the Iraq war, even though the local newpaper will print any offhand remark she makes on that topic. I care more about who she appointed to the committee to pick the dean of engineering, or about whether she wants to change the administrative status of sixth-year grad students. You’ll never read about those things in the newspaper.

So I’m pretty sure Summers wasn’t bounced because he dissed Cornel West and made some ill-considered remarks in a speech. Not having followed the details of Summers’ management of Harvard, I won’t pretend to know the detailed reasons for his ouster, or whether the Harvard Corporation showed good judgment in (apparently) deciding he should leave.

What is clear is that he is gone because the Corporation (what most other schools would call the Trustees or Regents) decided he should go. A faculty vote of no confidence last year had no real effect, and another one now would also not have mattered if the Corporation thought Summers was on the right track. If the faculty were involved in ousting Summers, their role was to convince the Corporation that Summers was doing a bad job.

Some commentators argue that the opinions of Harvard faculty shouldn’t matter. But even if Harvard faculty members know nothing about how Harvard should be run (which is pretty unlikely, if you ask me), it still matters what they think.

Consider a corporation run by a CEO who reports to a board of directors. If the majority of vice presidents think the CEO is doing a bad job, that should be a matter of concern for the board. They should talk to the CEO and the managers about what is happening. Then they should decide if people are unhappy because the CEO is making difficult but necessary decisions, or whether the CEO is just doing a bad job.

Maybe the CEO is doing an okay job at most things, but he seems to have a knack for angering and disappointing vice presidents. This is a problem for the company if it causes vice presidents to leave or makes it harder to recruit new ones. This problem is especially serious if other companies are eager to hire away vice presidents, and if the competence of the vice presidents is a big factor in the quality of the company’s output.

None of this depends on whether the vice presidents have the formal power to fire the CEO, or to do anything else for that matter. If employees make a difference in the company’s output and the labor market is competitive, the employees have power.

Which is why the call from some commentators to strip the faculty of their power is pointless. At most universities, the faculty have little or no formal power. All the Harvard Arts and Sciences faculty did was (a) pass non-binding resolutions, and (b) talk to people. To the extent they had power over the real decision makers, that power was granted not by Harvard but by the market. That is not something Harvard can change by amending its bylaws.


How Watermarks Fail

I wrote Wednesday about Randy Picker’s suggestion of using digital watermarks to embed users’ personal financial information into media files, to discourage users from sharing the files. Today, I want to talk more generally about watermarks and how they tend to fail.

First, some background. Watermarks are subtle signals embedded in the background of media files. They are supposed to be unobtrusive but easy to detect if you know where to look. Different media have different kinds of watermarks. In a photo, the watermark might be hidden in subtle patterns of shading. In music, it might be in a very soft background buzz, or a barely audible echo.

In many applications, a watermark must resist attempts by an adversary to remove it. For example, in Randy’s scheme, a user might want to remove the identifying watermark from a media file because he wants to share the file illegally, or because he doesn’t want his personal information exposed to cyber-intruders. It is often important to know how resistant a particular watermark is to removal. There has been plenty of research on this topic, from which we can draw lessons about how watermark removal tends to work.

One theme is the power of Rosetta Stone attacks. The original Rosetta Stone was a stone tablet with the same text written in three ancient languages. This gave scholars who understood one of the languages a big boost in deciphering another one that they didn’t understand. Similarly, watermarks tend to be defeated if an adversary can get his hands on a watermarked file, and the same file without the watermark. By comparing the two, the adversary can determine where the watermark lives, which is usually sufficient to remove the watermark from other files. Alex used this method in deciphering the MediaMax watermark (as described in our Sony CD DRM paper), and my colleagues and I used it also in analyzing the SDMI watermarks back in 2000.

Almost as powerful as a Rosetta Stone attack is a comparison attack, where the adversary does not have an unwatermarked file, but does have the same file with several different watermarks in it. Any place where two of the files differ is a place where watermark information lives. Given several marked files, an attacker can locate all or most of the places the watermark is hidden, which is again the first step in removing the watermark.

(In theory it might be possible to stop an adversary with access to a limited number of individually watermarked files from completely removing the watermark, if the watermark has lots of places to hide and is constructed cleverly. There is an interesting body of theory about how to do this and when it works. But in practice the assumptions underlying that theory rarely hold.)

Even if the adversary cannot get access to multiple versions of a file (so that Roseta Stone or comparison attacks are not possible), he can usually still defeat a watermark if he has access to a device that can detect watermarks. By reverse engineering the device, he can figure out where it is looking for the watermark, which again puts him in a position to remove it. (Even if he can’t dissect the device, he can use it as an oracle that tells him whether a particular file has a detectable watermark. Oracles are very helpful in attacking watermarks – Alex used one in his MediaMax watermark analysis, and my colleagues and I used one in our SDMI analysis.)

All of this helps us to understand where watermarks are likely to be effective and where they’re not. The best case for watermarking is where each file is published in a single version, with a watermark in a location that is not disclosed to the public and is not implemented in a device available to the public. This would hold true, for example, in a system that put a distinctive mark into all released versions of a file, and then looked for such watermarks in content broadcast on the radio or TV or downloaded from the net.

Not nearly as strong is a system where there is a single watermark per file, and consumer devices check for the mark – it is subject to reverse engineering and oracle attacks.

Weaker yet is a system where files are watermarked individually for each consumer – it is subject to comparison attacks.

Weakest of all is a system where files are watermarked individually for each consumer and everyone is told how to read the watermarks. Here the adversary can use comparison attacks, and reverse engineering is not even necessary because the inner working of the watermark detector are well known.

Alert readers will have noticed that all of the uses of watermarks for DRM (copy protection) seem to fall into the weak categories. That is because DRM applications require either that all devices check for the watermark – opening up reverse engineering and oracle attacks – or alternatively that a file be given separate watermarks for separate consumers – opening up comparison attacks. Watermarking has its uses, but it doesn’t seem well suited for DRM.


Mistrust-Based DRM

Randy Picker has an interesting post on the Chicago Law Faculty blog, describing what he calls “mistrust-based DRM”. The idea is that when an online music store gives you a song, it embeds into the song a watermark that contains your credit card number, or some other information that would let a (dishonest) person spend your money. This gives you an incentive not to distribute the song.

This is an instructive idea, but not a practical one.

In analyzing this idea, it’s helpful to divide it into two pieces: (1) embed a watermark that identifies the user, and (2) make that watermark a secret of the user and readable by the anyone who gets the file. Piece (1), taken alone, is a widely discussed DRM strategy which has not been used much in practice, for reasons I plan to discuss tomorrow. Today, I want to focus on the second piece.

Specifically, I want to compare two systems. In the more traditional system, the watermark is secret – it can be read only by the copyright owner or its agents – and users fear being sued for infringement if their files end up on P2P. In Randy’s system, the watermark is public – anybody can read it – and users fear being victimized by fraud if their files end up on P2P. I’ll call these two alternatives “secret-watermark” and “public-watermark”.

How do they compare? For starters, a secret watermark is much harder for an adversary to find and remove. If a watermark is public, everybody knows exactly where in the music it is stored. Common sense, and experience too, says that if you know where in a file information is stored, you can modify that part of the file and obliterate the information. But if the watermark is secret, then an adversary isn’t told where to look for it or how to change the file to remove it. Robustness of the watermark is an important issue that has been the downfall of past watermark systems.

A bigger problem with the public-watermark design, I think, are the forces unleashed when your design principle is to enable fraud. For example, the system will lose its force if unrelated anti-fraud measures become more effective, or if the financial system acts to protect users from fraud. Today, a consumer’s liability for fraudulent credit card transactions is capped at $50, and credit card companies often forgive even that $50. (You could use some other account information instead of the credit card number, but similar issues would still apply.) Copyright owners would be the only online merchants who wanted a higher level of fraud on the Net.

Worse yet, even law-abiding consumers would face a higher risk of fraud, because any loss or theft of their music or movie files would expose their financial information. Spyware programs could collect this information from users’ computers – and studies show that at least half of end-user PCs are infected with spyware. Law-abiding users would have a strong incentive to scrub the information out of their files, even if they had no intention of infringing. Alert anti-virus or anti-spyware vendors would be eager to provide this service.

Given the disadvantages of a public-watermark scheme, what are the arguments for it? Randy Picker argues that it gives end users an incentive to distrust fly-by-night purveyors of ripping software, worrying that they might steal the user’s information from the files and commit fraud. This isn’t entirely convincing: some such tools already contain heinous spyware that could cause users lots of harm, and reputable security suppliers are likely to provide watermark-scrubbing tools anyway. I think the threat of secret watermarks hidden in files, which fly-by-night vendors have no incentive to remove, would probably scare users enough.

On the whole, then, I think a secret-watermark scheme is better than a public-watermark one. But it should be noted that secret-watermark schemes themselves aren’t looking too good. They have mostly failed in the market, for reasons I’ll start digging into tomorrow.


Software Security: Creativity in a New Discipline

This is the last excerpt from my new book, Software Security: Building Security In. This might be a good time to buy the book.

Creativity in a New Discipline

We are experiencing a time of great creativity in computer security and must seize the opportunity presented by our current situation while we can. The diversity of backgrounds represented by today’s security practitioners may be a high-water mark. Consider that today’s security thought leaders were trained in fields as diverse as biostatistics, divinity, economics, and cognitive science, and thus bring with them interesting new perspectives on the security challenge. This leads to creative interplay in the field and has resulted in interesting progress, including the emergence of economic theories of security, an embrace of risk management, an emphasis on process-driven approaches (versus product sets), a shift toward software security, the rise of security engineering, and so on. As the worldwide security paradigm shift from guns, dogs, and concrete to networks, information systems, and computers continues unabated, we must leverage this time of creative diversity for all it’s worth.

A number of young researchers joined the computer security field in the mid-1990s, changing the focus of security research from spookware and national defense (think crypto, multilevel security, communications monitoring, and the like) to commercial systems and commerce. This movement away from military-oriented research was driven in part by the widespread public adoption of the Internet and the growing trend of e-commerce. With money at stake, security quickly became as relevant to business as it was to national defense. This influx of “new blood” shook up the scientific security research community and continues to have far-reaching effects that are only now affecting commercial security—the commercialization of firewalls, the rise of antivirus technology, and the adoption of modern security platforms, such as Java and .NET, were all predicted and spearheaded by new thinkers in the security research community.

Where Today’s Security People Come From

Only a handful of people working in computer security today started their careers in the field. In fact, academic programs expressly designed to train security practitioners are a recent phenomenon and remain rare.

Interestingly, it may be in this dearth of “qualified” people trained in security that a critical opportunity can be found. Though few practitioners have academic security training, they most assuredly do have academic training in some field of study. That means that as a collective, the computer security field is filled with diverse and interesting points of view. This is exactly the sort of Petri dish of ideas that led to the Renaissance at the end of the Dark Ages.

Diversity of ideas is healthy, and it lends a creativity and drive to the security field that we must take advantage of. A great example of this can be found in the new subfield of software security. Only five years ago the notion that bad software might be a major root cause of security issues was not common. Today, software security is the subject of keynote talks at the RSA security conference, and
we all seem to agree that we have a software problem to solve. This change was partially due to the involvement of programming languages people (once found only at obscure academic conferences like OOPSLA) in the security field. Such involvement resulted in the creation of modern languages like Java and .NET that include security models in their very design. When languages are declared “secure,” things get interesting! The evolutionary arms race between attackers and defenders jumps a level, new avenues for security design emerge, and dusty but thorny problems (think “buffer overflow”) become less relevant to the next generation of systems.

Where Tomorrow’s Security People Will Come From

These days, academic and professional training programs are being put in place to train the next generation of security professionals. Soon, standard curricula will be developed, and students will be required to understand the same core set of concepts. This will certainly help to solidify the field of computer security, but at the same time, there is a danger that generalization may lead to a homogenization of security. Instead of the creative soup afforded by a multiplicity of points of view spanning many fields, security runs the risk of becoming staid and static. If we are careful to avoid complete homogenization of the field, we can retain the benefits of diversity while building a solid academic discipline. One way to do this might be to encourage those students seeking computer security degrees to study widely in other supposedly unrelated disciplines as well. Another is to ensure that outside perspectives remain welcome in the field and are not dismissed out of hand. Computer security must remain an inclusive discipline in order to retain its creativity.

In any case, we must take advantage of the situation we find ourselves in now. Computer security is, in fact, experiencing an important rebirth, and now is the time to make great progress. We must pay close attention to different ideas, embrace change, and help security continue to evolve even as it begins to crystallize.


Software Security: A Case Study

Here is another excerpt from my new book, Software Security: Building Security In..

An Example: Java Card Security Testing

Doing effective security testing requires experience and knowledge. Examples and case studies like the one I present here are thus useful tools for understanding the approach.

In an effort to enhance payment cards with new functionality—such as the ability to provide secure cardholder identification or remember personal references—many credit-card companies are turning to multi-application smart cards. These cards use resident software applications to process and store thousands of times more information than traditional magnetic-stripe cards.

Security and fraud issues are critical concerns for the financial institutions and merchants spearheading smart-card adoption. By developing and deploying smart-card technology, credit-card companies provide important new tools in the effort to lower fraud and abuse. For instance, smart cards typically use a sophisticated crypto system to authenticate transactions and verify the identities of the cardholder and issuing bank. However, protecting against fraud and maintaining security and privacy are both very complex problems because of the rapidly evolving nature of smart-card technology.

The security community has been involved in security risk analysis and mitigation for Open Platform (now known as Global Platform, or GP) and Java Card since early 1997. Because product security is an essential aspect of credit-card companies’ brand protection regimen, companies like Visa and MasterCard spend plenty of time and effort on security testing and risk analysis. One central finding emphasizes the importance of testing particular vendor implementations according to our two testing categories: adherence to functional security design and proper behavior under particular attacks motivated by security risks.

The latter category, adversarial security testing (linked directly to risk analysis findings), ensures that cards can perform securely in the field even when under attack. Risk analysis results can be used to guide manual security testing. As an example, consider the risk that, as designed, the object-sharing mechanism in Java Card is complex and thus is likely to suffer from security-critical implementation errors on any given manufacturer’s card. Testing for this sort of risk involves creating and manipulating stored objects where sharing is involved. Given a technical description of this risk, building specific probing tests is possible.

Automating Security Testing

Over the years, Cigital has been involved in several projects that have identified architectural risks in the GP/Java Card platform, suggested several design improvements, and designed and built automated security tests for final products (each of which has multiple vendors).

Several years ago, we began developing an automated security test framework for GP cards built on Java Card 2.1.1 and based on extensive risk analysis results. The end result is a sophisticated test framework that runs with minimal human intervention and results in a qualitative security testing analysis of a sample smart card. This automated framework is now in use at MasterCard and the U.S. National Security Agency.

The first test set, the functional security test suite, directly probes low-level card security functionality. It includes automated testing of class codes, available commands, and crypto functionality. This test suite also actively probes for inappropriate card behavior of the sort that can lead to security compromise.

The second test set, the hostile applet test suite, is a sophisticated set of intentionally hostile Java Card applets designed to probe high-risk aspects of the GP on a Java Card implementation.

Results: Nonfunctional Security Testing Is Essential

Most cards tested with the automated test framework (but not all) pass all functional security tests, which we expect because smart-card vendors are diligent with functional testing (including security functionality). Because smart cards are complex embedded devices, vendors realize that exactly meeting functional requirements is an absolute necessity for customers to accept the cards. After all, they must perform properly worldwide.

However, every card submitted to the risk-based testing paradigm exhibited some manner of failure when tested with the hostile applet suite. Some failures pointed directly to critical security vulnerabilities on the card; others were less specific and required further exploration to determine the card’s true security posture.

As an example, consider that risk analysis of Java Card’s design documents indicates that proper implementation of atomic transaction processing is critical for maintaining a secure card. Java Card has the capability of defining transaction boundaries to ensure that if a transaction fails, data roll back to a pre-transaction state. In the event that transaction processing fails, transactions can go into any number of possible states, depending on what the applet was attempting. In the case of a stored-value card, bad transaction processing could allow an attacker to “print money” by forcing the card to roll back value counters while actually purchasing goods or services. This is called a “torn transaction” attack in credit-card risk lingo.

When creating risk-based tests to probe transaction processing, we directly exercised transaction-processing error handling by simulating an attacker attempting to violate a transaction—specifically, transactions were aborted or never committed, transaction buffers were completely filled, and transactions were nested (a no-no according to the Java Card specification). These tests were not based strictly on the card’s functionality—instead, security test engineers intentionally created them, thinking like an attacker given the results of a risk analysis.

Several real-world cards failed subsets of the transaction tests. The vulnerabilities discovered as a result of these tests would allow an attacker to terminate a transaction in a potentially advantageous manner—a critical test failure that wouldn’t have been uncovered under normal functional security testing. Fielding cards with these vulnerabilities would allow an attacker to execute successful attacks on live cards issued to the public. Because of proper risk-based security testing, the vendors were notified of the problems and corrected the code responsible before release.


Software Security: The Badness-ometer

Here is another excerpt from my new book, Software Security: Building Security In.

Application Security Tools: Good or Bad?

Application security testing products are being sold as a solution to the problem of insecure software. Unfortunately, these first-generation solutions are not all they are cracked up to be. They may help us diagnose, describe, and demonstrate the problem, but they do little to help us fix it.

Today’s application security products treat software applications as “black boxes” that are prone to misbehave and must be probed and prodded to prevent security disaster. Unfortunately, this approach is too simple.

Software testing requires planning and should be based on software requirements and the architecture of the code under test. You can’t “test quality in” by painstakingly finding and removing bugs once the code is done. The same goes for security; running a handful of canned tests that “simulate malicious hackers” by sending malformed input streams to a program will not work. Real attackers don’t simply “fuzz” a program with input to find problems. Attackers take software apart, determine how it works, and make it misbehave by doing what users are not supposed to do. The essence of the disconnect is that black box testing approaches, including application security testing tools, only scratch the surface of software in an outside→in fashion instead of digging into the guts of software and securing things from the inside.


That said, application security testing tools can tell you something about security—namely, that you’re in very deep trouble. That is, if your software fails any of the canned tests, you have some serious security work to do. The tools can help uncover known issues. But if you pass all the tests with flying colors, you know nothing more than that you passed a handful of tests with flying colors.

Put in more basic terms, application security testing tools are “badness-ometers,” as shown in the figure above. They provide a reading in a range from “deep trouble” to “who knows,” but they do not provide a reading into the “security” range at all. Most vulnerabilities that exist in the architecture and the code are beyond the reach of simple canned tests, so passing all the tests is not that reassuring. (Of course, knowing you’re in deep trouble can be helpful!)

The other major weakness with application security testing tools is that they focus only on input to an application provided over port 80. Understanding and testing a complex program by relying only on the protocol it uses to communicate provides a shallow analysis. Though many attacks do arrive via HTTP, this is only one category of security problem. First of all, input arrives to modern applications in many forms other than HTTP: consider SSL, environment variables, outside libraries, distributed components that communicate using other protocols, and so on. Beyond program input, software security must consider architectural soundness, data security, access control, software environment, and any number of other aspects, all of which are dependent on the application itself. There is no set of prefab tests that will probe every possible application in a meaningful way.

The only good use for application security tools is testing commercial off-the-shelf software. Simple dynamic checks set a reasonably low bar to hold vendors to. If software that is delivered to you fails to pass simple tests, you can either reject it out of hand or take steps to monitor its behavior.

In the final analysis, application security testing tools do provide a modicum of value. Organizations that are just beginning to think through software security issues can use them as badness-ometers to help determine how much trouble they are in. Results can alert all the interested parties to the presence of the problem and motivate some mitigation activity. However, you won’t get anything more than a rudimentary analysis with these tools. Fixing the problems they expose requires building better software to begin with—whether you created the software or not.


Software Security: The Trinity of Trouble

[Ed Felten says: Please welcome Gary McGraw as guest blogger for the next week. Gary is CTO at Cigital and co-author of two past books with me. He’s here to post excerpts from his new book, Software Security: Building Security In, which was released this week. The book offers practical advice about how to design and build secure software – a problem that is hugely important and often misunderstood. Now here’s Gary….]

The Trinity of Trouble: Why the Problem is Growing

Most modern computing systems are susceptible to software security problems, so why is software security a bigger problem now than in the past? Three trends—together making up the trinity of trouble—have a large influence on the growth and evolution of the problem.

Connectivity. The growing connectivity of computers through the Internet has increased both the number of attack vectors and the ease with which an attack can be made. This puts software at greater risk. More and more computers, ranging from home PCs to systems that control critical infrastructure, such as the supervisory control and data acquisition (SCADA) systems that run the power grid, are being connected to enterprise networks and to the Internet. Furthermore, people, businesses, and governments are increasingly dependent on network-enabled communication such as e-mail or Web pages provided by information systems. Things that used to happen offline now happen online. Unfortunately, as these systems are connected to the Internet, they become vulnerable to software-based attacks from distant sources. An attacker no longer needs physical access to a system to exploit vulnerable software; and today, software security problems can shut down banking services and airlines (as shown by the SQL Slammer worm of January 2003).

Because access through a network does not require human intervention, launching automated attacks is easy. The ubiquity of networking means that there are more software systems to attack, more attacks, and greater risks from poor software security practices than in the past. We’re really only now beginning to cope with the ten-year-old attack paradigm that results from poor coding and design. Ubiquitous networking and attacks directly related to distributed computation remain rare (though the network itself is the primary vector for getting to and exploiting poor coding and design problems). This will change for the worse over time. Because the Internet is everywhere, the attackers are now at your virtual doorstep.

To make matters worse, large enterprises have caught two bugs: Web Services and its closely aligned Service Oriented Architecture (SOA). Even though SOA is certainly a fad driven by clever marketing, it represents a succinct way to talk about what many security professionals have always known to be true: Legacy applications that were never intended to be internetworked are becoming inter-networked and published as services.

Common platforms being integrated into megasolutions include SAP, PeopleSoft, Oracle, Informatica, Maestro, and so on (not to mention more modern J2EE and NET apps), COBOL, and other ancient mainframe platforms. Many of these applications and legacy systems don’t support common toolkits like SSL, standard plug-ins for authentication/authorization in a connected situation, or even simple cipher use. They don’t have the builtin capability to hook into directory services, which most large shops use for authentication and authorization. Middleware vendors pledge they can completely carve out the complexity of integration and provide seamless connectivity, but even though they provide connectivity (through JCA, WBI, or whatever), the authentication and application-level protocols don’t align.

Thus, middleware integration in reality reduces to something ad hoc like cross-enterprise FTP between applications. What’s worse is that lines of business often fear tight integration with better tools (because they lack skills, project budget, or faith in their infrastructure team), so they end up using middleware to FTP and drop data globs that have to be mopped up and transmogrified into load files or other application input. Because of this issue, legacy product integrations often suffer from two huge security problems:
1. Exclusive reliance on host-to-host authentication with weak passwords
2. Looming data compliance implications having to do with user privacy (because unencrypted transport of data over middleware and the middleware’s implementation for failover and load balancing means that queue cache files get stashed all over the place in plain text)

Current trends in enterprise architecture make connectivity problems more problematic than ever before.

Extensibility. A second trend negatively affecting software security is the degree to which systems have become extensible. An extensible system accepts updates or extensions, sometimes referred to as mobile code so that the functionality of the system can be evolved in an incremental fashion. For example, the plug-in architecture of Web browsers makes it easy to install viewer extensions for new document types as needed. Today’s operating systems support extensibility through dynamically loadable device drivers and modules. Today’s applications, such as word processors, e-mail clients, spreadsheets, and Web browsers, support extensibility through scripting, controls, components, and applets. The advent of Web Services and SOA, which are built entirely from extensible systems such as J2EE and .NET, brings explicit extensibility to the forefront.

From an economic standpoint, extensible systems are attractive because they provide flexible interfaces that can be adapted through new components. In today’s marketplace, it is crucial that software be deployed as rapidly as possible in order to gain market share. Yet the marketplace also demands that applications provide new features with each release. An extensible architecture makes it easy to satisfy both demands by allowing the base application code to be shipped early, with later feature extensions shipped as needed.

Unfortunately, the very nature of extensible systems makes it hard to prevent software vulnerabilities from slipping in as unwanted extensions. Advanced languages and platforms including Sun Microsystems’ Java and Microsoft’s .NET Framework are making extensibility commonplace.

Complexity. A third trend impacting software security is the unbridled growth in the size and complexity of modern information systems, especially software systems. A desktop system running Windows XP and associated applications depends on the proper functioning of the kernel as well as the applications to ensure that vulnerabilities cannot compromise the system. However, Windows XP itself consists of at least forty million lines of code, and end-user applications are becoming equally, if not more, complex. When systems become this large, bugs cannot be avoided.

The figure above shows how the complexity of Windows (measured in lines of code) has grown over the years. The point of the graph is not to emphasize the numbers themselves, but rather the growth rate over time. In practice, the defect rate tends to go up as the square of code size. Other factors that significantly affect complexity include whether the code is tightly integrated, the overlay of patches and other post-deployment fixes, and critical architectural issues.

The complexity problem is exacerbated by the use of unsafe programming languages (e.g., C and C++) that do not protect against simple kinds of attacks, such as buffer overflows. In theory, we could analyze and prove that a small program was free of problems, but this task is impossible for even the simplest desktop systems today, much less the enterprise-wide systems used by businesses or governments.

Of course, Windows is not alone. Almost all code bases tend to grow over time. During the last three years, I have made an informal survey of thousands of developers. With few exceptions (on the order of 1% of sample size), developers overwhelmingly report that their groups intend to produce more code, not less, as time goes by. Ironically, these same developers also report that they intend to produce fewer bugs even as they produce more code. The unfortunate reality is that “more lines, more bugs” is the rule of thumb that tends to be borne out in practice (and in science, as the next section shows). Developers are an optimistic lot.

The propensity for software systems to grow very large quickly is just as apparent in open source systems as it is in Windows. The problem is, of course, that more code results in more defects and, in turn, more security risk.


Sony CD DRM Paper Released

Today Alex and I released our paper about the Sony CD DRM episode. This is the full, extended version of the paper, with a bunch of new material that hasn’t been published or posted before.

As an experiment, we posted draft sections of the paper here and asked readers for comments and feedback. The experiment was a success, giving us lots of good comments and suggestions that helped us improve the paper. Several reader-commenters are thanked in the paper’s acknowledgments section.

We also asked readers to help suggest a title for the paper. That didn’t work out so well – some suggestions were entertaining, but none were really practical. Perhaps a title of the sort we wanted doesn’t exist.

Enjoy the paper, and thanks for your help.

[UPDATE (Feb. 21): If you don’t like PDFs, you can now read the paper in your browser, thanks to an HTML+images version created by Jesse Weinstein.]


Secure Flight Mothballed

Secure Flight, the planned next-generation system for screening airline passengers, has been mothballed by the Transportation Security Administration, according to an AP story by Leslie Miller. TSA chief Kip Hawley cited security concerns and questions about the program’s overall direction.

Last year I served on the Secure Flight Working Group, a committee of outside technology and privacy experts asked by the TSA to give feedback on Secure Flight. After hearing about plans for Secure Flight, I was convinced that TSA didn’t have a clear idea of what the program was supposed to be doing or how it would work. This is essentially what later government studies of the program found. Here’s the AP story:

Nearly four years and $200 million after the program was put into operation, Hawley said last month that the agency hadn’t yet determined precisely how it would work.

Government auditors gave the project failing grades – twice – and rebuked its authors for secretly obtaining personal information about airline passengers.

The sad part of this is that Secure Flight seems to have started out as a simpler program that would have made sense to deploy.

Today, airlines are given a no-fly list and a watch-list, which they are asked to check against their passenger lists. There are obvious security drawbacks to distributing the lists to airlines – a malicious airline employee with access to the lists could leak them to the bad guys. The 9/11 Commission recommended keeping the lists within the government, and having the government check passengers’ names against the lists.

A program designed to do just that would have been a good idea. There would still be design issues to work out. For example, false matches are now handled by airline ticket agents, but that function would probably have to moved into the government too, which would raise some logistical issues. There would be privacy worries, but they could be handled with good design and oversight.

Instead of sticking to this more modest plan, Secure Flight became a vehicle for pie-in-the-sky plans about data mining and automatic identification of terrorists from consumer databases. As the program’s goals grew more ambitious and collided with practical design and deployment challenges, the program lost focus and seemed to have a different rationale and plan from one month to the next.

What happens now is predictable. The program will officially die but will actually be reincarnated with a new name. Congress has directed TSA to implement a program of this general type, so TSA really has no choice but to try again. Let’s hope that this time they make the hard choices they avoided last time, and end up with a simpler program that solves the easier problems first.

(Fellow Working Group member Lauren Gelman offers has a similar take on this story. Another member, Bruce Schneier, has also blogged extensively about Secure Flight.)


Quality of Service: A Quality Argument?

One of the standard arguments one hears against network neutrality rules is that network providers need to provide Quality of Service (QoS) guarantees to certain kinds of traffic, such as video. If QoS is necessary, the argument goes, and if net neutrality rules would hamper QoS by requiring all traffic to be treated the same, then net neutrality rules must be harmful. Today, I want to unpack this argument and see how it holds up in light of computer science research and engineering experience.

First, I need to make clear that guaranteeing QoS for an application means more than just giving it lots of bandwidth or prioritizing its traffic above other applications. Those things might be helpful, but they’re not QoS (or at least not the kind I’m talking about today). What QoS mechanisms (try to) do is to make specific performance guarantees to an app over a short window of time.

An example may clarify this point. If you’re loading a web page, and your network connection hiccups so that you get no traffic for (say) half a second, you may notice a short pause but it won’t be a big deal. But if you’re having a voice conversation with somebody, a half-second gap will be very annoying. Web browsing needs decent bandwidth on average, but voice conversations needs better protection against short delays. That protection is QoS.

Careful readers will protest at this point that a good browsing experience depends on more than just average bandwidth. A half-second hiccup might not be a big problem, but a ten-minute pause would be too much, even if performance is really snappy afterward. The difference between voice conversations and browsing is one of degree – voice conversations want guarantees over fractions of seconds, and browsing wants them over fractions of minutes.

The reason we don’t need special QoS mechanisms for browsing is that the broadband Internet already provides performance that is almost always steady enough over the time intervals that matter for browsing.

Sometimes, too, there are simple tricks that can turn an app that cares about short delays into one that cares only about longer delays. For example, watching prerecorded audio or video streams doesn’t need QoS, because you can use buffering. If you’re watching a video, you can download every frame ten seconds before you’re going to watch it; then a hiccup of a few seconds won’t be a problem. This is why streaming audio and video work perfectly well today (when there is enough average bandwidth).

There are two other important cases where QoS isn’t needed. First, if an app needs higher average speed than the Net can provide, than QoS won’t help it – QoS makes the Net’s speed steadier but not faster. Second – and less obvious – if an app needs much less average speed than the Net can provide, then QoS might also be unnecessary. If speed doesn’t drop entirely to zero but fluctuates, with peaks and valleys, then even the valleys may be high enough to give the app what it needs. This is starting to happen for voice conversations – Skype and other VoIP systems seem to work pretty well without any special QoS support in the network.

We can’t say that QoS is never needed, but experience does teach that it’s easy, especially for non-experts, to overestimate the importance of QoS. That’s why I’m not convinced – though I could be, with more evidence – that QoS is a strong argument against net neutrality rules.