ACTION:371, Re: de-identification text for Wednesday's call

Hi:

I am happy that the tone in this thread has regained a good tone, and there have been many substantive comments on the thread.

Here are three questions I have as we prepare for today's call:

(1) What disagreements remain on the proposed normative text?  I have seen Roy's email that focuses on not re-identifying a person.  Are there are other proposed changes to the normative text?  Our call today can examine these options.

(2) My sense is that there is interest in having Shane, Dan, and perhaps others work more on non-normative text including examples, hopefully for next week's call.  Our call today can identify others interested in such an effort, and identify any other proposed changes to non-normative text.

(3) I have a question based in part on the presentation last week on financial auditing and accounting.  The trend, according to the Deloitte and Touche experts, is for revenue related to online ads to become "material" for a larger number of companies, including many publicly-traded companies that have to follow Sarbanes-Oxley rules to retain data for seven years.  To respond to audit requirements, the trend similarly appears to be to retain detailed log files about what ads were served when and where.  Absent some idea from the group, it is difficult to see how DNT can substantially reduce these retention requirements for audit purposes.

My question is how this often-required retention affects the discussion between Dan and Shane about administrative vs. technical controls for de-identification?  I understand Dan to be supporting an approach that has a first round of significant technical controls, with administrative controls as secondary at most.  I understand Shane to be supporting the possibility of all-administrative controls.  (My apologies if this simplification is incorrect.)

In a world where the auditors and accountants will often require retention for seven years, does this affect our view of the role of administrative controls?  I would think that administrative controls are the main way that auditors would retain access, while others cannot.

One answer is that administrative controls may be appropriate for auditors and accountants, and that is a reason for strong access controls and other measures to wall them off from the rest of an organization.  After all, auditors and accountants have a long history of confidentiality.  Meanwhile, our discussion of de-identification applies to everyone else.

Another answer is that our de-identification definitions are important for future practices and codes of conduct in the industry generally.  I think that is true.  On this view, our discussion is helping create a precedent for what counts as effective de-identification more generally.  In that view, our discussion of the relative role of administrative vs. technical controls is clearly important in many contexts.

I welcome thoughts on the role of the de-identification definition in the context of possible retention for permitted uses such as audit and accounting.

Peter



Professor Peter P. Swire
C. William O'Neill Professor of Law
    Ohio State University
240.994.4142
www.peterswire.net

From: Dan Auerbach <dan@eff.org<mailto:dan@eff.org>>
Date: Wednesday, April 3, 2013 3:42 AM
To: "public-tracking@w3.org<mailto:public-tracking@w3.org>" <public-tracking@w3.org<mailto:public-tracking@w3.org>>
Subject: Re: de-identification text for Wednesday's call
Resent-From: <public-tracking@w3.org<mailto:public-tracking@w3.org>>
Resent-Date: Wednesday, April 3, 2013 3:43 AM

On 04/02/2013 10:15 PM, Shane Wiley wrote:
Dan,

Thank you for being open to further discussion (hopefully not considered ‘simply arguing’).

I can’t speak to your experience with companies you’ve worked for in the past but I can say that a company that is truly committed to Accountability (as a Privacy concept) takes technical, operational, and administrative controls very seriously (documentation, design reviews, testing, auditing, etc.).  It appears you’re seeking a way to capture bad actors for not living up to their promises (say they de-identify and then due to a poorly designed de-identification process or operational controls they allow re-identification to occur – or perhaps more likely that data reserved for a Permitted Use is used for a non-permitted use) – is that correct?
Permitted uses are limited in the scope of data collection and retention based on their function. On the other hand, since out of scope for DNT, de-identified data can be retained for dozens or hundreds of years. As such I think it is a much more dangerous avenue for abuse than a permitted use. If by a "bad actor" you mean someone who doesn't thoroughly de-identify data, then yes, this is my concern. But I don't think it takes a malicious actor to become convinced that a certain de-identification process is good enough, when it is not adequate. One litmus test would be: if a court order forced you to turn over lots of data to law enforcement (one of a myriad of concerns a user could have when setting DNT:1), could law enforcement plausibly infer information about individuals? If yes, I think that the company did not do its duty to de-identify based on the normative text.

Perhaps privacy penetration testing might offer a way forward, if in fact we disagree about the re-identification or attribute disclosure risk. But we may not disagree about the risk, just the acceptable level and type of risk. I'm not sure. Here's a concrete scenario to help sort this out: suppose at the end of your proposed de-identification process, you let Arvind Narayanan loose on your data. Suppose he is able to poke holes in your methodology and infer information about individuals. Would this require you to change your data practices? Could you still be considered a good actor, even if Arvind could plausibly poke these holes, and yet you aren't making any changes to the de-identification process in response?



If that is correct, I don’t have an easy answer for you as this same issue exists for almost all company privacy promises that involve internal management of data.  I would recommend we take the broader topic of ‘external scrutiny of internal data practices’ to a different thread/working group.  It’s my assumption that if a W3C DNT standard emerges that does invite voluntary implementation by industry, that trade associations will step in to include this in their Codes of Conduct and regular audits of member compliance.  I don’t see an immediate solution that would give EFF access to every online company’s backend systems to self-review their practices for compliance.

I share your desire to provide meaningful guidance to implementers of de-identification and do this in a way as to not limit innovation in this area and not lock companies into solutions that harm their ability to operate as a business.  I don’t believe your non-normative text strikes this balance and offered to come back to the group next week with a few alternate examples that better find this balance.  If you’re open to that approach, we can open an action item for me to deliver next week.
I'm certainly happy to read your example text, and I hope we can come to consensus about it. But I suspect we have a deep disagreement, and think we need to have evidence-based discussions going forward. Perhaps, as you suggest, we could have competing text in the draft document, to be resolved later. We could also try to tackle this head on, if you and I (and anyone interested) separately try to come to some agreement about the text and report back to the group. Happy to take suggestions about the best way forward, I'm really not sure.


Thank you,
Shane

From: Dan Auerbach [mailto:dan@eff.org]
Sent: Tuesday, April 02, 2013 3:59 PM
To: public-tracking@w3.org<mailto:public-tracking@w3.org>
Subject: Re: de-identification text for Wednesday's call

Shane,

Thanks for your response, and I suppose we can talk more tomorrow, though we should probably try to identify a path forward as opposed to simply arguing. For example, I would welcome any empirical investigation into the efficacy of various approaches, including operational ones.

I think the spectrum is wider than you suggest and ranges from no change at all to deletion of data. In practice I think deleting data is easy and for many actors the most straight forward way to comply with DNT. My position is far less aggressive than insisting on deleting data. On what I would consider to be the "lighter" end of the spectrum, there are solutions like yours that espouse an undefined mix of technical data manipulation measures and operational controls. Having worked in industry, I'm very skeptical of the value of operational controls, and think that in practice there will be very little accountability. If you'd like to provide more detailed background about the operational controls you have in mind or empirical evidence of the efficacy of such controls, I am certainly willing to listen.

I agree that we shouldn't be too prescriptive about particular methods of de-identification, but I think it's important for the standard to come a clear conclusion about the place in the spectrum that counts as good enough. I don't think that saying, "there's a spectrum, and anything even on the lighter side is OK because companies will be held accountable" is the right answer. My view is that we need to be specific enough to define the contours of what is acceptable, without being too prescriptive, and that's the balance that I hoped to strike with my normative and non-normative text.

On 04/02/2013 02:36 PM, Shane Wiley wrote:
Dan – my apologies – I used those terms to represent one end of a spectrum.  Let’s step away from those particular terms to represent the most conservative side of the spectrum and instead speak to the range of possible points of application of consumer protection with respect to de-identification.

On one side of the spectrum we have a possible solution that is completely based on administrative and operational controls (as you pointed out earlier) and on the other side of the spectrum is a possible solution that completely relies on technology alone (“and provides for leeway as anonymization technology improves”).  In the middle is a solution that is a mix and allows organizations the latitude to devise a solution that matches their ability to mitigate risk and maintain some value in data – while not being overly prescriptive on the specific technology employed.  All points on the spectrum focus on the desired outcome – that de-identified data not later be re-identified.

While we both agree on the outcome (I believe), we disagree on the best route to get there.  I would suggest we move away from prescriptive elements in the text and provide organizations the flexibility to innovate and build solutions they believe meets the goal.  As long as a company is stating they will meet the goal, you have all of the necessary commitment to hold them accountable to that outcome.

Evidence:  I believe the perceived harm is that de-identified data can later be re-identified, fair?  If yes, then any solution in the spectrum serves that purposes but forces companies most likely to the middle or more conservative side of the spectrum to mitigate their risks without being overly prescriptive.

Industry implementation of a standard that is too weak will be worse for users than having no standard at all, since it will provide only the guise of protection:  Compared to today’s world, even this “in the middle” solution to de-identification GREATLY moves the needle in the correct direction – I hope you agree.
As a user, I'd prefer having no DNT at all to a de-identification approach that I would consider too soft. At least with no DNT, I have no false expectation that my data isn't being collected and retained.


  Will there be bad actors or companies that overstate their ability to maintain a de-identified state and fail?  Absolutely.  This approach ensures they will be appropriately held accountable and at the same time allow good actors an attractive (voluntary) option to enhance consumer privacy.

Your concerns:  I do not mean to dismiss your concerns and hopefully have done a better job this time of fairly representing a shared goal and spectrum of options to achieve that goal.  I believe our best path forward is to agree on the goal and avoid prescriptive requirements on exactly how to achieve it.  While we each have our own views into what the rest of the Working Group holds out as priorities (which can only be proven through a vote which the W3C likes to avoid), I hope this compromise position both represents a meaningful advance for consumer protection for consumer advocates and provides enough flexibility for voluntary adoption by industry.

- Shane

From: Dan Auerbach [mailto:dan@eff.org]
Sent: Tuesday, April 02, 2013 1:39 PM
To: public-tracking@w3.org<mailto:public-tracking@w3.org>
Subject: Re: de-identification text for Wednesday's call

Shane,

Labeling my view as "totalitarian" or "absolutist" is inaccurate and not appreciated. My approach allows lots of leeway as anonymization technology improves, and takes into account that perfect anonymization is impossible. It seems to me that you plan to label anything "totalitarian" that suggests that keeping raw logs is not a viable approach in general. To be clear, I don't even want to suggest that keeping raw logs is *always* prohibited, just that it should be in the normal case where records have relatively high entropy. I'm also happy to discuss examples in great detail, but in my experience you have been unwilling to engage in discussing the nitty gritty details of what anonymization should entail.

On 04/02/2013 01:09 PM, Shane Wiley wrote:
Dan,

As with HIPPA, I believe differentiated treatment of internal and external datasets is appropriate as this changes the risk profile of re-identification – again, the root of the conversation being a “risk-based” approach versus a totalitarian approach as you suggest.  My solution meets the perceived consumer harm in this case
Do you have evidence for this claim?




– yours of course does as well but goes far too far over the top of what is actually needed.  If your goal is to create a compromise end-point that will likely be implemented by industry then my recommended approach gets us there.
Any compromise must provide meaningful protection to users. Industry implementation of a standard that is too weak will be worse for users than having no standard at all, since it will provide only the guise of protection.



  If you’d like to instead stand by absolutist approaches, that is of course your prerogative and we’ll have those removed through the standard W3C process.  I’m simply trying to save everyone some time and get to a meaningful outcome quickly.
Glibly dismissing my concerns is not a way to gain allies, or move the W3C process forward. I don't think your view is as universally shared as you seem to think that it is.





- Shane

From: Dan Auerbach [mailto:dan@eff.org]
Sent: Tuesday, April 02, 2013 12:21 PM
To: public-tracking@w3.org<mailto:public-tracking@w3.org>
Subject: Re: de-identification text for Wednesday's call

Shane,

Why hash at all in this case? If you are relying on operational and administrative controls, you might as well just pledge not to look up the cookie when you receive it. If you are rotating (and discarding) salts frequently, then it will have a positive effect, but otherwise I don't think hashing provides any benefit here.

But this is an aside to our main disagreement about the larger issue about the role that operational and administrative controls should play. I agree that they should play a role, but only after de-identification of data has been achieved. If the result of a DNT:1 request is business as usual, with minor scrubbing and the caveat that only 4000 engineers at a large corporation get default access to a specially marked database instead of 10000, then that will not be a successful standard. (Of course I welcome more detailed information about operational and administrative controls.)

One last point I wanted to make is that of course the data sets I mentioned refer to public data. We don't have access to internal corporate data sets. There are laws in place to protect the pilfering of that data, so of course no-one is going to steal data then publish an academic paper about it, effectively painting a big target on themselves for federal prosecutors and corporate legal teams. In light of this, the right empirical question to ask is: of large publicly available data sets that contain user data and are somewhat akin to log data, how often are there successful re-identification or attribute disclosure attacks? Can you point to any public data sets where such an attack has not been found?

If your argument is instead that public data should be treated differently from non-public data, then I'd suggest that this is out of scope for the DNT conversation. DNT is about giving users the choice to opt out of tracking by companies, which must entail meaningfully curbing data collection and retention by that company, not merely a request that a company not make public its collected data. (Indeed, in addition to being an excessively weak demand by the user, this would in some cases be a vacuous request, since making that information public is already prohibited by law.) The de-identification question exists within the scope of what the companies themselves can do with the data -- is the data de-identified with respect to the entity that collected the data?

Best,
Dan

On 04/02/2013 11:03 AM, Shane Wiley wrote:
Dan,

Once the one-way hash is applied (and other elements of record appropriately cleansed) the data is moved to a system that is not allowed to be accessed externally.  Its these operational and administrative controls that are essential to ensure de-identified data is not re-identified at some later time.  I believe you’re looking only at the technical merits which is only seeing a small portion of the overall solution.

- Shane

From: Dan Auerbach [mailto:dan@eff.org]
Sent: Tuesday, April 02, 2013 10:59 AM
To: public-tracking@w3.org<mailto:public-tracking@w3.org>
Subject: Re: de-identification text for Wednesday's call

On 04/02/2013 08:50 AM, Shane Wiley wrote:
once the one-way hash function has been applied the data is never again able to be accessed in real-time to modify the user’s experience.
I think I'm confused, can you explain this more? How is this possible? If you are just hashing a cookie string, your web server receives a request that includes a cookie string, you hash that cookie string (which is in incredibly fast operation), match the hashed cookie against the stored data, and return personalized results.

Or are you salting the hash differently for every request, or combining the cookie with an ephemeral piece of data (the timestamp) before hashing and then throwing away the timestamp?

Thanks for clarifying, apologies if I'm just being dense.

Dan






--

Dan Auerbach

Staff Technologist

Electronic Frontier Foundation

dan@eff.org<mailto:dan@eff.org>

415 436 9333 x134






--

Dan Auerbach

Staff Technologist

Electronic Frontier Foundation

dan@eff.org<mailto:dan@eff.org>

415 436 9333 x134





--

Dan Auerbach

Staff Technologist

Electronic Frontier Foundation

dan@eff.org<mailto:dan@eff.org>

415 436 9333 x134




--

Dan Auerbach

Staff Technologist

Electronic Frontier Foundation

dan@eff.org<mailto:dan@eff.org>

415 436 9333 x134



--
Dan Auerbach
Staff Technologist
Electronic Frontier Foundation
dan@eff.org<mailto:dan@eff.org>
415 436 9333 x134