March 19, 2024

Archives for May 2011

Studying the Frequency of Redaction Failures in PACER

Since we launched RECAP a couple of years ago, one of our top concerns has been privacy. The federal judiciary’s PACER system offers the public online access to hundreds of millions of court records. The judiciary’s rules require each party in a case to redact certain types of information from documents they submit, but unfortunately litigants and their counsel don’t always comply with these rules. Three years ago, Carl Malamud did a groundbreaking audit of PACER documents and found more than 1600 cases in which litigants submitted documents with unredacted Social Security numbers. My recent research has focused on a different problem: cases where parties tried to redact sensitive information but the redactions failed for technical reasons. This problem occasionally pops up in news stories, but as far as I know, no one has conducted a systematic study.

To understand the problem, it helps to know a little bit about how computers represent graphics. The simplest image formats are bitmap or raster formats. These represent an image as an array of pixels, with each pixel having a color represented by a numeric value. The PDF format uses a different approach, known as vector graphics, that represent an image as a series of drawing commands: lines, rectangles, lines of text, and so forth.

Vector graphics have important advantages. Vector-based formats “scale up” gracefully, in contrast to the raster images that look “blocky” at high resolutions. Vector graphics also do a better job of preserving a document’s structure. For example, text in a PDF is represented by a sequence of explicit text-drawing commands, which is why you can cut and paste text from a PDF document, but not from a raster format like PNG.

But vector-based formats also have an important disadvantage: they may contain more information than is visible to the naked eye. Raster images have a “what you see is what you get” quality—changing all the pixels in a particular region to black destroys the information that was previously in that part of the image. But a vector-based image can have multiple “layers.” There might be a command to draw some text followed by a command to draw a black rectangle over the text. The image might look like it’s been redacted, but the text is still “under” the box. And often extracting that information is a simple matter of cutting and pasting.

So how many PACER documents have this problem? We’re in a good position to study this question because we have a large collection of PACER documents—1.8 million of them when I started my research last year. I wrote software to detect redaction rectangles—it turns out these are relatively easy to recognize based on their color, shape, and the specific commands used to draw them. Out of 1.8 million PACER documents, there were approximately 2000 documents with redaction rectangles. (There were also about 3500 documents that were redacted by replacing text by strings of Xes, I also excluded documents that were redacted by Carl Malamud before he donated them to our archive.)

Next, my software checked to see if these redaction rectangles overlapped with text. My software identified a few hundred documents that appeared to have text under redaction rectangles, and examining them by hand revealed 194 documents with failed redactions. The majority of the documents (about 130) appear be from commercial litigation, in which parties have unsuccessfully attempted to redact trade secrets such as sales figures and confidential product information. Other improperly redacted documents contain sensitive medical information, addresses, and dates of birth. Still others contain the names of witnesses, jurors, plaintiffs, and one minor.

Implications

PACER reportedly contains about 500 million documents. We don’t have a random sample of PACER documents, so we should be careful about trying to extrapolate to the entire PACER corpus. Still, it’s safe to say there are thousands, and probably tens of thousands, of documents in PACER whose authors made unsuccessful attempts to conceal information.

It’s also important to note that my software may not be detecting every instance of redaction failures. If a PDF was created by scanning in a paper document (as opposed to generated directly from a word processor), then it probably won’t have a “text layer.” My software doesn’t detect redaction failures in this type of document. This means that there may be more than 194 failed redactions among the 1.8 million documents I studied.

A few weeks ago I wrote a letter to Judge Lee Rosenthal, chair of the federal judiciary’s Committee on Rules of Practice and Procedure, explaining this problem. In that letter I recommend that the courts themselves use software like mine to automatically scan PACER documents for this type of problem. In addition to scanning the documents they already have, the courts should make it a standard part of the process for filing new documents with the courts. This would allow the courts to catch these problems before the documents are made available to the public on the PACER website.

My code is available here. It’s experimental research code, not a finished product. We’re releasing it into the public domain using the CC0 license; this should make it easy for federal and state officials to adapt it for their own use. Court administrators who are interested in adapting the code for their own use are especially encouraged to contact me for advice and assistance. The code relies heavily on the CAM::PDF Perl library, and I’m indebted to Chris Dolan for his patient answers to my many dumb questions.

Getting Redaction Right

So what should litigants do to avoid this problem? The National Security Agency has a good primer on secure redaction. The approach they recommend—completely deleting sensitive information in the original word processing document, replacing it with innocuous filler (such as strings of XXes) as needed, and then converting it to a PDF document, is the safest approach. The NSA primer also explains how to check for other potentially sensitive information that might be hidden in a document’s metadata.

Of course, there may be cases where this approach isn’t feasible because a litigant doesn’t have the original word processing document or doesn’t want the document’s layout to be changed by the redaction process. Adobe Acrobat’s redaction tool has worked correctly when we’ve used it, and Adobe probably has the expertise to do it correctly. There may be other tools that work correctly, but we haven’t had an opportunity to experiment with them so we can’t say which ones they might be.

Regardless of the tool used, it’s a good idea to take the redacted document and double-check that the information was removed. An easy way to do this is to simply cut and paste the “redacted” content into another document. If the redaction succeeded, no text should be transferred. This method will catch most, but not all, redaction failures. A more rigorous check is to remove the redaction rectangles from the document and manually observe what’s underneath them. One of the scripts I’m releasing today, called remove_rectangles.pl, does just that. In its current form, it’s probably not user-friendly enough for non-programmers to use, but it would be relatively straightforward for someone (perhaps Adobe or the courts) to build a user-friendly version that ordinary users could use to verify that the document they just attempted to redact actually got redacted.

One approach we don’t endorse is printing the document out, redacting it with a black marker, and then re-scanning it to PDF format. Although this may succeed in removing the sensitive information, we don’t recommend this approach because it effectively converts the document into a raster-based image, destroying useful information in the process. For example, it will no longer be possible to cut and paste (non-redacted) text from a document that has been redacted in this way.

Bad redactions are not a new problem, but they are taking on a new urgency as PACER documents become increasingly available on the web. Correct redaction is not difficult, but it does require both knowledge and care by those who are submitting the documents. The courts have several important roles they should play: educating attorneys about their redaction responsibilities, providing them with software tools that make it easy for them to comply, and monitoring submitted documents to verify that the rules are being followed.

This research was made possible with the financial support of Carl Malamud’s organization, Public.Resource.Org.

"You Might Also Like:" Privacy Risks of Collaborative Filtering

Ann Kilzer, Arvind Narayanan, Ed Felten, Vitaly Shmatikov, and I have released a new research paper detailing the privacy risks posed by collaborative filtering recommender systems. To examine the risk, we use public data available from Hunch, LibraryThing, Last.fm, and Amazon in addition to evaluating a synthetic system using data from the Netflix Prize dataset. The results demonstrate that temporal changes in recommendations can reveal purchases or other transactions of individual users.

To help users find items of interest, sites routinely recommend items similar to a given item. For example, product pages on Amazon contain a “Customers Who Bought This Item Also Bought” list. These recommendations are typically public, and they are the product of patterns learned from all users of the system. If customers often purchase both item A and item B, a collaborative filtering system will judge them to be highly similar. Most sites generate ordered lists of similar items for any given item, but some also provide numeric similarity scores.

Although item similarity is only indirectly related to individual transactions, we determined that temporal changes in item similarity lists or scores can reveal details of those transactions. If you’re a Mozart fan and you listen to a Justin Bieber song, this choice increases the perceived similarity between Justin Bieber and Mozart. Because similarity lists and scores are based on perceived similarity, your action may result in changes to these scores or lists.

Suppose that an attacker knows some of your past purchases on a site: for example, past item reviews, social networking profiles, or real-world interactions are a rich source of information. New purchases will affect the perceived similarity between the new items and your past purchases, possibility causing visible changes to the recommendations provided for your previously purchased items. We demonstrate that an attacker can leverage these observable changes to infer your purchases. Among other things, these attacks are complicated by the fact that multiple users simultaneously interact with a system and updates are not immediate following a transaction.

To evaluate our attacks, we use data from Hunch, LibraryThing, Last.fm, and Amazon. Our goal is not to claim privacy flaws in these specific sites (in fact, we often use data voluntarily disclosed by their users to verify our inferences), but to demonstrate the general feasibility of inferring individual transactions from the outputs of collaborative filtering systems. Among their many differences, these sites vary dramatically in the information that they reveal. For example, Hunch reveals raw item-to-item correlation scores, but Amazon reveals only lists of similar items. In addition, we examine a simulated system created using the Netflix Prize dataset. Our paper outlines the experimental results.

While inference of a Justin Bieber interest may be innocuous, inferences could expose anything from dissatisfaction with a job to health issues. Our attacks assume that a victim reveals certain past transactions, but users may publicly reveal certain transactions while preferring to keep others private. Ultimately, users are best equipped to determine which transactions would be embarrassing or otherwise problematic. We demonstrate that the public outputs of recommender systems can reveal transactions without user knowledge or consent.

Unfortunately, existing privacy technologies appear inadequate here, failing to simultaneously guarantee acceptable recommendation quality and user privacy. Mitigation strategies are a rich area for future work, and we hope to work towards solutions with others in the community.

Worth noting is that this work suggests a risk posed by any feature that adapts in response to potentially sensitive user actions. Unless sites explicitly consider the data exposed, such features may inadvertently leak details of these underlying actions.

Our paper contains additional details. This work was presented earlier today at the 2011 IEEE Symposium on Security and Privacy. Arvind has also blogged about this work.

Web Tracking and User Privacy Workshop: Test Cases for Privacy on the Web

This guest post is from Nick Doty, of the W3C and UC Berkeley School of Information. As a companion post to my summary of the position papers submitted for last month’s W3C Do-Not-Track Workshop, hosted by CITP, Nick goes deeper into the substance and interaction during the workshop.

The level of interest and participation in last month’s Workshop on Web Tracking and User Privacy — about a hundred attendees spanning multiple countries, dozens of companies, a wide variety of backgrounds — confirms the broad interest in Do Not Track. The relatively straightforward technical approach with a catchy name has led to, in the US, proposed legislation at both the state and federal level and specific mention by the Federal Trade Commission (it was nice to have Ed Felten back from DC representing his new employer at the workshop), and comparatively rapid deployment of competing proposals by browser vendors. Still, one might be surprised that so many players are devoting such engineering resources to a relatively narrow goal: building technical means that allow users to avoid tracking across the Web for the purpose of compiling behavioral profiles for targeted advertising.

In fact, Do Not Track (in all its variations and competing proposals) is the latest test case for how new online technologies will address privacy issues. What mix of minimization techniques (where one might classify Microsoft’s Tracking Protection block lists) versus preference expression and use limitation (like a Do Not Track header) will best protect privacy and allow for innovation? Can parties agree on a machine-readable expression of privacy preferences (as has been heavily debated in P3P, GeoPriv and other standards work), and if so, how will terms be defined and compliance monitored and enforced? Many attendees were at the workshop not just to address this particular privacy problem — ubiquitous invisible tracking of Web requests to build behavioral profiles — but to grab a seat at the table where the future of how privacy is handled on the Web may be decided. The W3C, for its part, expects to start an Interest Group to monitor privacy on the Web and spin out specific work as new privacy issues inevitably arise, in addition to considering a Working Group to address this particular topic (more below). The Internet Engineering Task Force (IETF) is exploring a Privacy Directorate to provide guidance on privacy considerations across specs.

At a higher level, this debate presents a test case for the process of building consensus and developing standards around technologies like tracking protection or Do Not Track that have inspired controversy. What body (or rather, combination of bodies) can legitimately define preference expressions that must operate at multiple levels in the Web stack, not to mention serve the diverse needs of individuals and entities across the globe? Can the same organization that defines the technical design also negotiate semantic agreement between very diverse groups on the meaning of “tracking”? Is this an appropriate role for technical standards bodies to assume? To what extent can technical groups work with policymakers to build solutions that can be enforced by self-regulatory or governmental players?

Discussion at the recent workshop confirmed many of these complexities: though the agenda was organized to roughly separate user experience, technical granularity, enforcement and standardization, overlap was common and inevitable. Proposals for an “ack” or response header brought up questions of whether the opportunity to disclaim following the preference would prevent legal enforcement; whether not having such a response would leave users confused about when they had opted back in; and how granular such header responses should be. In defining first vs. third party tracking, user expectations, current Web business models and even the same-origin security policy could point the group in different directions.

We did see some moments of consensus. There was general agreement that while user interface issues were key to privacy, trying to standardize those elements was probably counterproductive but providing guidance could help significantly. Regarding the scope of “tracking”, the group was roughly evenly divided on what they would most prefer: a broad definition (any logging), a narrow definition (online behavioral advertising profiling only) or something in between (where tracking is more than OBA but excludes things like analytics or fraud protection, as in the proposal from the Center for Democracy and Technology). But in a “hum” to see which proposals workshop attendees opposed (“non-starters”) no one objected to starting with a CDT-style middle ground — a rather shocking level of agreement to end two days chock full of debate.

For tech policy nerds, then, this intimate workshop about a couple of narrow technical proposals was heady stuff. And the points of agreement suggest that real interoperable progress on tracking protection — the kind that will help the average end user’s privacy — is on the way. For the W3C, this will certainly be a topic of discussion at the ongoing meeting in Bilbao, and we’re beginning detailed conversations about the scope and milestones for a Working Group to undertake technical standards work.

Thanks again to Princeton/CITP for hosting the event, and to Thomas and Lorrie for organizing it: bringing together this diverse group of people on short notice was a real challenge, and it paid off for all of us. If you’d like to see more primary materials: minutes from the workshop (including presentations and discussions) are available, as are the position papers and slides. And the W3C will post a workshop report with a more detailed summary very soon.

Overstock's $1M Challenge

As reported in Fast Company, RichRelevance and Overstock.com teamed up to offer up to a $1,000,000 prize for improving “its recommendation engine by 10 percent or more.”

If You Liked Netflix, You Might Also Like Overstock
When I first read a summary of this contest, it appeared they were following in Netflix’s footsteps right down to releasing user data sans names. This did not end well for Netflix’s users or for Netflix. Narayanan and Shmatikov were able to re-identify Netflix users using the contest dataset, and their research contributed greatly to Ohm’s work on de-anonimization. After running the contest a second time, Netflix terminated it early in the face of FTC attention and a lawsuit that they settled out of court.

This time, Overstock is providing “synthetic data” to contest entrants, then testing submitted algorithms against unreleased real data. Tag line: “If you can’t bring the data to the code, bring the code to the data.” Hmm. An interesting idea, but short on a few details around the sharp edges that jump out as highest concern. I look forward to getting the time to play with the system and dataset. What is good news is seeing companies recognize privacy concerns and respond with something interesting and new. That is, at least, a move in the right direction.

Place your bets now on which happens first: a contest winner with a 10% boost to sales, or researchers finding ways to re-identify at least 10% of the data?

Debugging Legislation: PROTECT IP

There’s more than a hint of theatrics in the draft PROTECT IP bill (pdf, via dontcensortheinternet ) that has emerged as son-of-COICA, starting with the ungainly acronym of a name. Given its roots in the entertainment industry, that low drama comes as no surprise. Each section name is worse than the last: “Eliminating the Financial Incentive to Steal Intellectual Property Online” (Sec. 4) gives way to “Voluntary action for Taking Action Against Websites Stealing American Intellectual Property” (Sec. 5).

Techdirt gives a good overview of the bill, so I’ll just pick some details:

  • Infringing activities. In defining “infringing activities,” the draft explicitly includes circumvention devices (“offering goods or services in violation of section 1201 of title 17”), as well as copyright infringement and trademark counterfeiting. Yet that definition also brackets the possibility of “no [substantial/significant] use other than ….” Substantial could incorporate the “merely capable of substantial non-infringing use” test of Betamax.
  • Blocking non-domestic sites. Sec. 3 gives the Attorney General a right of action over “nondomestic domain names”, including the right to demand remedies from (A) domain name system server operators, (B) financial transaction providers, (C), Internet advertising services, and (D) “an interactive computer service (def. from 230(f)) shall take technically feasible and reasonable measures … to remove or disable access to the Internet site associated with the domain name set forth in the order, or a hypertext link to such Internet site.”
  • Private right of action. Sec. 3 and Sec. 4 appear to be near duplicates (I say appear, because unlike computer code, we don’t have a macro function to replace the plaintiff, so the whole text is repeated with no diff), replacing nondomestic domain with “domain” and permitting private plaintiffs — “a holder of an intellectual property right harmed by the activities of an Internet site dedicated to infringing activities occurring on that Internet site.” Oddly, the statute doesn’t say the simpler “one whose rights are infringed,” so the definition must be broader. Could a movie studio claim to be hurt by the infringement of others’ rights, or MPAA enforce on behalf of all its members? Sec. 4 is missing (d)(2)(D)
  • WHOIS. The “applicable publicly accessible database of registrations” gets a new role as source of notice for the domain registrant, “to the extent such addresses are reasonably available.” (c)(1)
  • Remedies. The bill specifies injunctive relief only, not money damages, but threat of an injunction can be backed by the unspecified threat of contempt for violating one.
  • Voluntary action. Finally the bill leaves room for “voluntary action” by financial transaction providers and advertising services, immunizing them from liability to anyone if they choose to stop providing service, notwithstanding any agreements to the contrary. This provision jeopardizes the security of online businesses, making them unable to contract for financial services against the possibility that someone will wrongly accuse them of infringement. 5(a) We’ve already seen that it takes little to convince service providers to kick users off, in the face of pressure short of full legal process (see everyone vs Wikileaks, Facebook booting activists, and numerous misfired DMCA takedowns); this provision insulates that insecurity further.

In short, rather than “protecting” intellectual and creative industry, this bill would make it less secure, giving the U.S. a competitive disadvantage in online business. (Sorry, Harlan, that we still can’t debug the US Code as true code.)